Eval bug: Gemma 3 extremly slow prompt processing when using quantized kv cache. #12352

Bearsaerker · 2025-03-12T11:47:37Z

Name and Version

./build/bin/llama-cli --version
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: yes
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
version: 0 (unknown)
built with cc (GCC) 13.3.1 20240611 (Red Hat 13.3.1-2) for x86_64-redhat-linux

(newest b4876 version)

Operating systems

Linux

GGML backends

CUDA

Hardware

Ryzen 3900x + rtx 3060 12gb

Models

Gemma-3-12b_Q5_K_M

Problem description & steps to reproduce

Prompt eval time is way slower when using quantized kv cache than standard kv cache. Also I see that the cpu is used when the quantized kv cache is turned on. So I believe that the kv cache is not properly processed by the gpu if the quantized kv cache is provided

First Bad Commit

No response

Relevant log output

# Unquantized kv cache:

./build/bin/llama-server -m '/home/luis/Downloads/llama.cpp-b4876/models/gemma-3-12b-it-Q5_K_M.gguf'  --n-gpu-layers -1 --batch_size 1024 --flash-attn -c 4000 --port 7777 -t 8 -ngl 99
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: yes
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
build: 0 (unknown) with cc (GCC) 13.3.1 20240611 (Red Hat 13.3.1-2) for x86_64-redhat-linux
system info: n_threads = 8, n_threads_batch = 8, total_threads = 24

system_info: n_threads = 8 (n_threads_batch = 8) / 24 | CUDA : ARCHS = 860 | FORCE_CUBLAS = 1 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

main: HTTP server is listening, hostname: 127.0.0.1, port: 7777, http threads: 23
main: loading model
srv    load_model: loading model '/home/luis/Downloads/llama.cpp-b4876/models/gemma-3-12b-it-Q5_K_M.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3060) - 10456 MiB free
llama_model_loader: loaded meta data with 34 key-value pairs and 626 tensors from /home/luis/Downloads/llama.cpp-b4876/models/gemma-3-12b-it-Q5_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = gemma3
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Gemma 3
llama_model_loader: - kv   3:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   4:                         general.size_label str              = 12B
llama_model_loader: - kv   5:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv   6:                      gemma3.context_length u32              = 131072
llama_model_loader: - kv   7:                    gemma3.embedding_length u32              = 3840
llama_model_loader: - kv   8:                         gemma3.block_count u32              = 48
llama_model_loader: - kv   9:                 gemma3.feed_forward_length u32              = 15360
llama_model_loader: - kv  10:                gemma3.attention.head_count u32              = 16
llama_model_loader: - kv  11:    gemma3.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  12:                gemma3.attention.key_length u32              = 256
llama_model_loader: - kv  13:              gemma3.attention.value_length u32              = 256
llama_model_loader: - kv  14:                      gemma3.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  15:            gemma3.attention.sliding_window u32              = 1024
llama_model_loader: - kv  16:             gemma3.attention.head_count_kv u32              = 8
llama_model_loader: - kv  17:                   gemma3.rope.scaling.type str              = linear
llama_model_loader: - kv  18:                 gemma3.rope.scaling.factor f32              = 8.000000
llama_model_loader: - kv  19:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  20:                         tokenizer.ggml.pre str              = default
llama_model_loader: - kv  21:                      tokenizer.ggml.tokens arr[str,262208]  = ["<pad>", "<eos>", "<bos>", "<unk>", ...
llama_model_loader: - kv  22:                      tokenizer.ggml.scores arr[f32,262208]  = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  23:                  tokenizer.ggml.token_type arr[i32,262208]  = [3, 3, 3, 3, 3, 4, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv  24:                tokenizer.ggml.bos_token_id u32              = 2
llama_model_loader: - kv  25:                tokenizer.ggml.eos_token_id u32              = 106
llama_model_loader: - kv  26:            tokenizer.ggml.unknown_token_id u32              = 3
llama_model_loader: - kv  27:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  28:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  29:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  30:                    tokenizer.chat_template str              = {{ bos_token }}\n{%- if messages[0]['r...
llama_model_loader: - kv  31:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  32:               general.quantization_version u32              = 2
llama_model_loader: - kv  33:                          general.file_type u32              = 17
llama_model_loader: - type  f32:  289 tensors
llama_model_loader: - type q5_K:  288 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q5_K - Medium
print_info: file size   = 7.86 GiB (5.74 BPW) 
load: special tokens cache size = 6415
load: token to piece cache size = 1.9446 MB
print_info: arch             = gemma3
print_info: vocab_only       = 0
print_info: n_ctx_train      = 131072
print_info: n_embd           = 3840
print_info: n_layer          = 48
print_info: n_head           = 16
print_info: n_head_kv        = 8
print_info: n_rot            = 256
print_info: n_swa            = 1024
print_info: n_embd_head_k    = 256
print_info: n_embd_head_v    = 256
print_info: n_gqa            = 2
print_info: n_embd_k_gqa     = 2048
print_info: n_embd_v_gqa     = 2048
print_info: f_norm_eps       = 0.0e+00
print_info: f_norm_rms_eps   = 1.0e-06
print_info: f_clamp_kqv      = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale    = 0.0e+00
print_info: f_attn_scale     = 6.2e-02
print_info: n_ff             = 15360
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 2
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000.0
print_info: freq_scale_train = 0.125
print_info: n_ctx_orig_yarn  = 131072
print_info: rope_finetuned   = unknown
print_info: ssm_d_conv       = 0
print_info: ssm_d_inner      = 0
print_info: ssm_d_state      = 0
print_info: ssm_dt_rank      = 0
print_info: ssm_dt_b_c_rms   = 0
print_info: model type       = 12B
print_info: model params     = 11.77 B
print_info: general.name     = Gemma 3
print_info: vocab type       = SPM
print_info: n_vocab          = 262208
print_info: n_merges         = 0
print_info: BOS token        = 2 '<bos>'
print_info: EOS token        = 106 '<end_of_turn>'
print_info: EOT token        = 106 '<end_of_turn>'
print_info: UNK token        = 3 '<unk>'
print_info: PAD token        = 0 '<pad>'
print_info: LF token         = 248 '<0x0A>'
print_info: EOG token        = 106 '<end_of_turn>'
print_info: max token length = 48
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 48 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 49/49 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   787.69 MiB
load_tensors:        CUDA0 model buffer size =  8047.63 MiB
.....................................................................................
llama_init_from_model: n_seq_max     = 1
llama_init_from_model: n_ctx         = 4096
llama_init_from_model: n_ctx_per_seq = 4096
llama_init_from_model: n_batch       = 1024
llama_init_from_model: n_ubatch      = 512
llama_init_from_model: flash_attn    = 1
llama_init_from_model: freq_base     = 1000000.0
llama_init_from_model: freq_scale    = 0.125
llama_init_from_model: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_kv_cache_init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 48, can_shift = 1
llama_kv_cache_init:      CUDA0 KV buffer size =  1536.00 MiB
llama_init_from_model: KV self size  = 1536.00 MiB, K (f16):  768.00 MiB, V (f16):  768.00 MiB
llama_init_from_model:  CUDA_Host  output buffer size =     1.00 MiB
llama_init_from_model:      CUDA0 compute buffer size =   519.62 MiB
llama_init_from_model:  CUDA_Host compute buffer size =    23.51 MiB
llama_init_from_model: graph nodes  = 1737
llama_init_from_model: graph splits = 2
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv          init: initializing slots, n_slots = 1
slot         init: id  0 | task -1 | new slot n_ctx_slot = 4096
main: model loaded
main: chat template, chat_template: {{ bos_token }}
{%- if messages[0]['role'] == 'system' -%}
    {%- if messages[0]['content'] is string -%}
        {%- set first_user_prefix = messages[0]['content'] + '

' -%}
    {%- else -%}
        {%- set first_user_prefix = messages[0]['content'][0]['text'] + '

' -%}
    {%- endif -%}
    {%- set loop_messages = messages[1:] -%}
{%- else -%}
    {%- set first_user_prefix = "" -%}
    {%- set loop_messages = messages -%}
{%- endif -%}
{%- for message in loop_messages -%}
    {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
        {{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
    {%- endif -%}
    {%- if (message['role'] == 'assistant') -%}
        {%- set role = "model" -%}
    {%- else -%}
        {%- set role = message['role'] -%}
    {%- endif -%}
    {{ '<start_of_turn>' + role + '
' + (first_user_prefix if loop.first else "") }}
    {%- if message['content'] is string -%}
        {{ message['content'] | trim }}
    {%- elif message['content'] is iterable -%}
        {%- for item in message['content'] -%}
            {%- if item['type'] == 'image' -%}
                {{ '<start_of_image>' }}
            {%- elif item['type'] == 'text' -%}
                {{ item['text'] | trim }}
            {%- endif -%}
        {%- endfor -%}
    {%- else -%}
        {{ raise_exception("Invalid content type") }}
    {%- endif -%}
    {{ '<end_of_turn>
' }}
{%- endfor -%}
{%- if add_generation_prompt -%}
    {{'<start_of_turn>model
'}}
{%- endif -%}
, example_format: '<start_of_turn>user
You are a helpful assistant

Hello<end_of_turn>
<start_of_turn>model
Hi there<end_of_turn>
<start_of_turn>user
How are you?<end_of_turn>
<start_of_turn>model
'
main: server is listening on http://127.0.0.1:7777 - starting the main loop
srv  update_slots: all slots are idle
srv  params_from_: Chat format: Content-only
slot launch_slot_: id  0 | task 0 | processing task
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, n_prompt_tokens = 2505
slot update_slots: id  0 | task 0 | kv cache rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 1024, n_tokens = 1024, progress = 0.408782
slot update_slots: id  0 | task 0 | kv cache rm [1024, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 2048, n_tokens = 1024, progress = 0.817565
slot update_slots: id  0 | task 0 | kv cache rm [2048, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 2505, n_tokens = 457, progress = 1.000000
slot update_slots: id  0 | task 0 | prompt done, n_past = 2505, n_tokens = 457
slot      release: id  0 | task 0 | stop processing: n_past = 3199, truncated = 0
slot print_timing: id  0 | task 0 | 
prompt eval time =    2697.39 ms /  2505 tokens (    1.08 ms per token,   928.67 tokens per second)
       eval time =   24911.73 ms /   695 tokens (   35.84 ms per token,    27.90 tokens per second)
      total time =   27609.12 ms /  3200 tokens
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200


# Quantized kv cache:
 ./build/bin/llama-server -m '/home/luis/Downloads/llama.cpp-b4876/models/gemma-3-12b-it-Q5_K_M.gguf'  --n-gpu-layers -1 --cache-type-k q8_0 --cache-type-v q8_0 --batch_size 1024 --flash-attn -c 4000 --port 7777 -t 8 -ngl 99
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: yes
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
build: 0 (unknown) with cc (GCC) 13.3.1 20240611 (Red Hat 13.3.1-2) for x86_64-redhat-linux
system info: n_threads = 8, n_threads_batch = 8, total_threads = 24

system_info: n_threads = 8 (n_threads_batch = 8) / 24 | CUDA : ARCHS = 860 | FORCE_CUBLAS = 1 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

main: HTTP server is listening, hostname: 127.0.0.1, port: 7777, http threads: 23
main: loading model
srv    load_model: loading model '/home/luis/Downloads/llama.cpp-b4876/models/gemma-3-12b-it-Q5_K_M.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3060) - 10500 MiB free
llama_model_loader: loaded meta data with 34 key-value pairs and 626 tensors from /home/luis/Downloads/llama.cpp-b4876/models/gemma-3-12b-it-Q5_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = gemma3
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Gemma 3
llama_model_loader: - kv   3:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   4:                         general.size_label str              = 12B
llama_model_loader: - kv   5:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv   6:                      gemma3.context_length u32              = 131072
llama_model_loader: - kv   7:                    gemma3.embedding_length u32              = 3840
llama_model_loader: - kv   8:                         gemma3.block_count u32              = 48
llama_model_loader: - kv   9:                 gemma3.feed_forward_length u32              = 15360
llama_model_loader: - kv  10:                gemma3.attention.head_count u32              = 16
llama_model_loader: - kv  11:    gemma3.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  12:                gemma3.attention.key_length u32              = 256
llama_model_loader: - kv  13:              gemma3.attention.value_length u32              = 256
llama_model_loader: - kv  14:                      gemma3.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  15:            gemma3.attention.sliding_window u32              = 1024
llama_model_loader: - kv  16:             gemma3.attention.head_count_kv u32              = 8
llama_model_loader: - kv  17:                   gemma3.rope.scaling.type str              = linear
llama_model_loader: - kv  18:                 gemma3.rope.scaling.factor f32              = 8.000000
llama_model_loader: - kv  19:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  20:                         tokenizer.ggml.pre str              = default
llama_model_loader: - kv  21:                      tokenizer.ggml.tokens arr[str,262208]  = ["<pad>", "<eos>", "<bos>", "<unk>", ...
llama_model_loader: - kv  22:                      tokenizer.ggml.scores arr[f32,262208]  = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  23:                  tokenizer.ggml.token_type arr[i32,262208]  = [3, 3, 3, 3, 3, 4, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv  24:                tokenizer.ggml.bos_token_id u32              = 2
llama_model_loader: - kv  25:                tokenizer.ggml.eos_token_id u32              = 106
llama_model_loader: - kv  26:            tokenizer.ggml.unknown_token_id u32              = 3
llama_model_loader: - kv  27:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  28:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  29:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  30:                    tokenizer.chat_template str              = {{ bos_token }}\n{%- if messages[0]['r...
llama_model_loader: - kv  31:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  32:               general.quantization_version u32              = 2
llama_model_loader: - kv  33:                          general.file_type u32              = 17
llama_model_loader: - type  f32:  289 tensors
llama_model_loader: - type q5_K:  288 tensors
llama_model_loader: - type q6_K:   49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q5_K - Medium
print_info: file size   = 7.86 GiB (5.74 BPW) 
load: special tokens cache size = 6415
load: token to piece cache size = 1.9446 MB
print_info: arch             = gemma3
print_info: vocab_only       = 0
print_info: n_ctx_train      = 131072
print_info: n_embd           = 3840
print_info: n_layer          = 48
print_info: n_head           = 16
print_info: n_head_kv        = 8
print_info: n_rot            = 256
print_info: n_swa            = 1024
print_info: n_embd_head_k    = 256
print_info: n_embd_head_v    = 256
print_info: n_gqa            = 2
print_info: n_embd_k_gqa     = 2048
print_info: n_embd_v_gqa     = 2048
print_info: f_norm_eps       = 0.0e+00
print_info: f_norm_rms_eps   = 1.0e-06
print_info: f_clamp_kqv      = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale    = 0.0e+00
print_info: f_attn_scale     = 6.2e-02
print_info: n_ff             = 15360
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 2
print_info: rope scaling     = linear
print_info: freq_base_train  = 1000000.0
print_info: freq_scale_train = 0.125
print_info: n_ctx_orig_yarn  = 131072
print_info: rope_finetuned   = unknown
print_info: ssm_d_conv       = 0
print_info: ssm_d_inner      = 0
print_info: ssm_d_state      = 0
print_info: ssm_dt_rank      = 0
print_info: ssm_dt_b_c_rms   = 0
print_info: model type       = 12B
print_info: model params     = 11.77 B
print_info: general.name     = Gemma 3
print_info: vocab type       = SPM
print_info: n_vocab          = 262208
print_info: n_merges         = 0
print_info: BOS token        = 2 '<bos>'
print_info: EOS token        = 106 '<end_of_turn>'
print_info: EOT token        = 106 '<end_of_turn>'
print_info: UNK token        = 3 '<unk>'
print_info: PAD token        = 0 '<pad>'
print_info: LF token         = 248 '<0x0A>'
print_info: EOG token        = 106 '<end_of_turn>'
print_info: max token length = 48
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 48 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 49/49 layers to GPU
load_tensors:   CPU_Mapped model buffer size =   787.69 MiB
load_tensors:        CUDA0 model buffer size =  8047.63 MiB
.....................................................................................
llama_init_from_model: n_seq_max     = 1
llama_init_from_model: n_ctx         = 4096
llama_init_from_model: n_ctx_per_seq = 4096
llama_init_from_model: n_batch       = 1024
llama_init_from_model: n_ubatch      = 512
llama_init_from_model: flash_attn    = 1
llama_init_from_model: freq_base     = 1000000.0
llama_init_from_model: freq_scale    = 0.125
llama_init_from_model: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_kv_cache_init: kv_size = 4096, offload = 1, type_k = 'q8_0', type_v = 'q8_0', n_layer = 48, can_shift = 1
llama_kv_cache_init:      CUDA0 KV buffer size =   816.00 MiB
llama_init_from_model: KV self size  =  816.00 MiB, K (q8_0):  408.00 MiB, V (q8_0):  408.00 MiB
llama_init_from_model:  CUDA_Host  output buffer size =     1.00 MiB
llama_init_from_model:      CUDA0 compute buffer size =   519.62 MiB
llama_init_from_model:  CUDA_Host compute buffer size =    45.01 MiB
llama_init_from_model: graph nodes  = 1737
llama_init_from_model: graph splits = 98
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv          init: initializing slots, n_slots = 1
slot         init: id  0 | task -1 | new slot n_ctx_slot = 4096
main: model loaded
main: chat template, chat_template: {{ bos_token }}
{%- if messages[0]['role'] == 'system' -%}
    {%- if messages[0]['content'] is string -%}
        {%- set first_user_prefix = messages[0]['content'] + '

' -%}
    {%- else -%}
        {%- set first_user_prefix = messages[0]['content'][0]['text'] + '

' -%}
    {%- endif -%}
    {%- set loop_messages = messages[1:] -%}
{%- else -%}
    {%- set first_user_prefix = "" -%}
    {%- set loop_messages = messages -%}
{%- endif -%}
{%- for message in loop_messages -%}
    {%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
        {{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
    {%- endif -%}
    {%- if (message['role'] == 'assistant') -%}
        {%- set role = "model" -%}
    {%- else -%}
        {%- set role = message['role'] -%}
    {%- endif -%}
    {{ '<start_of_turn>' + role + '
' + (first_user_prefix if loop.first else "") }}
    {%- if message['content'] is string -%}
        {{ message['content'] | trim }}
    {%- elif message['content'] is iterable -%}
        {%- for item in message['content'] -%}
            {%- if item['type'] == 'image' -%}
                {{ '<start_of_image>' }}
            {%- elif item['type'] == 'text' -%}
                {{ item['text'] | trim }}
            {%- endif -%}
        {%- endfor -%}
    {%- else -%}
        {{ raise_exception("Invalid content type") }}
    {%- endif -%}
    {{ '<end_of_turn>
' }}
{%- endfor -%}
{%- if add_generation_prompt -%}
    {{'<start_of_turn>model
'}}
{%- endif -%}
, example_format: '<start_of_turn>user
You are a helpful assistant

Hello<end_of_turn>
<start_of_turn>model
Hi there<end_of_turn>
<start_of_turn>user
How are you?<end_of_turn>
<start_of_turn>model
'
main: server is listening on http://127.0.0.1:7777 - starting the main loop
srv  update_slots: all slots are idle
srv  params_from_: Chat format: Content-only
slot launch_slot_: id  0 | task 0 | processing task
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, n_prompt_tokens = 2505
slot update_slots: id  0 | task 0 | kv cache rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 1024, n_tokens = 1024, progress = 0.408782
slot update_slots: id  0 | task 0 | kv cache rm [1024, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 2048, n_tokens = 1024, progress = 0.817565
slot update_slots: id  0 | task 0 | kv cache rm [2048, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_past = 2505, n_tokens = 457, progress = 1.000000
slot update_slots: id  0 | task 0 | prompt done, n_past = 2505, n_tokens = 457
slot      release: id  0 | task 0 | stop processing: n_past = 3209, truncated = 0
slot print_timing: id  0 | task 0 | 
prompt eval time =   23150.35 ms /  2505 tokens (    9.24 ms per token,   108.21 tokens per second)
       eval time =   78076.05 ms /   705 tokens (  110.75 ms per token,     9.03 tokens per second)
      total time =  101226.41 ms /  3210 tokens
srv  update_slots: all slots are idle
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200

Andybui1012 · 2025-03-12T12:05:32Z

The same bug also happens with RekaAI/reka-flash-3.

szghds · 2025-03-12T12:05:54Z

Also encountered the same problem.

slaren · 2025-03-12T12:51:11Z

ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: yes

Try a build without this option. If you don't remember enabling it, delete the build directory first and reconfigure cmake.

Bearsaerker · 2025-03-12T13:03:14Z

Unfortunately did not change a thing

./bin/llama-server -m '/home/luis/Downloads/llama.cpp-b4876/models/gemma-3-12b-it-Q5_K_M.gguf' --n-gpu-layers -1 --cache-type-k q8_0 --cache-type-v q8_0 --batch_size 1024 --flash-attn -c 4000 --port 7777 -t 8 -ngl 99
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GeForce RTX 3060, compute capability 8.6, VMM: yes
build: 0 (unknown) with cc (GCC) 13.3.1 20240611 (Red Hat 13.3.1-2) for x86_64-redhat-linux
system info: n_threads = 8, n_threads_batch = 8, total_threads = 24

main: HTTP server is listening, hostname: 127.0.0.1, port: 7777, http threads: 23
main: loading model
srv load_model: loading model '/home/luis/Downloads/llama.cpp-b4876/models/gemma-3-12b-it-Q5_K_M.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 3060) - 10444 MiB free
llama_model_loader: loaded meta data with 34 key-value pairs and 626 tensors from /home/luis/Downloads/llama.cpp-b4876/models/gemma-3-12b-it-Q5_K_M.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv 0: general.architecture str = gemma3
llama_model_loader: - kv 1: general.type str = model
llama_model_loader: - kv 2: general.name str = Gemma 3
llama_model_loader: - kv 3: general.quantized_by str = Unsloth
llama_model_loader: - kv 4: general.size_label str = 12B
llama_model_loader: - kv 5: general.repo_url str = https://huggingface.co/unsloth
llama_model_loader: - kv 6: gemma3.context_length u32 = 131072
llama_model_loader: - kv 7: gemma3.embedding_length u32 = 3840
llama_model_loader: - kv 8: gemma3.block_count u32 = 48
llama_model_loader: - kv 9: gemma3.feed_forward_length u32 = 15360
llama_model_loader: - kv 10: gemma3.attention.head_count u32 = 16
llama_model_loader: - kv 11: gemma3.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 12: gemma3.attention.key_length u32 = 256
llama_model_loader: - kv 13: gemma3.attention.value_length u32 = 256
llama_model_loader: - kv 14: gemma3.rope.freq_base f32 = 1000000.000000
llama_model_loader: - kv 15: gemma3.attention.sliding_window u32 = 1024
llama_model_loader: - kv 16: gemma3.attention.head_count_kv u32 = 8
llama_model_loader: - kv 17: gemma3.rope.scaling.type str = linear
llama_model_loader: - kv 18: gemma3.rope.scaling.factor f32 = 8.000000
llama_model_loader: - kv 19: tokenizer.ggml.model str = llama
llama_model_loader: - kv 20: tokenizer.ggml.pre str = default
llama_model_loader: - kv 21: tokenizer.ggml.tokens arr[str,262208] = ["", "", "", "", ...
llama_model_loader: - kv 22: tokenizer.ggml.scores arr[f32,262208] = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv 23: tokenizer.ggml.token_type arr[i32,262208] = [3, 3, 3, 3, 3, 4, 3, 3, 3, 3, 3, 3, ...
llama_model_loader: - kv 24: tokenizer.ggml.bos_token_id u32 = 2
llama_model_loader: - kv 25: tokenizer.ggml.eos_token_id u32 = 106
llama_model_loader: - kv 26: tokenizer.ggml.unknown_token_id u32 = 3
llama_model_loader: - kv 27: tokenizer.ggml.padding_token_id u32 = 0
llama_model_loader: - kv 28: tokenizer.ggml.add_bos_token bool = true
llama_model_loader: - kv 29: tokenizer.ggml.add_eos_token bool = false
llama_model_loader: - kv 30: tokenizer.chat_template str = {{ bos_token }}\n{%- if messages[0]['r...
llama_model_loader: - kv 31: tokenizer.ggml.add_space_prefix bool = false
llama_model_loader: - kv 32: general.quantization_version u32 = 2
llama_model_loader: - kv 33: general.file_type u32 = 17
llama_model_loader: - type f32: 289 tensors
llama_model_loader: - type q5_K: 288 tensors
llama_model_loader: - type q6_K: 49 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type = Q5_K - Medium
print_info: file size = 7.86 GiB (5.74 BPW)
load: special tokens cache size = 6415
load: token to piece cache size = 1.9446 MB
print_info: arch = gemma3
print_info: vocab_only = 0
print_info: n_ctx_train = 131072
print_info: n_embd = 3840
print_info: n_layer = 48
print_info: n_head = 16
print_info: n_head_kv = 8
print_info: n_rot = 256
print_info: n_swa = 1024
print_info: n_embd_head_k = 256
print_info: n_embd_head_v = 256
print_info: n_gqa = 2
print_info: n_embd_k_gqa = 2048
print_info: n_embd_v_gqa = 2048
print_info: f_norm_eps = 0.0e+00
print_info: f_norm_rms_eps = 1.0e-06
print_info: f_clamp_kqv = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale = 0.0e+00
print_info: f_attn_scale = 6.2e-02
print_info: n_ff = 15360
print_info: n_expert = 0
print_info: n_expert_used = 0
print_info: causal attn = 1
print_info: pooling type = 0
print_info: rope type = 2
print_info: rope scaling = linear
print_info: freq_base_train = 1000000.0
print_info: freq_scale_train = 0.125
print_info: n_ctx_orig_yarn = 131072
print_info: rope_finetuned = unknown
print_info: ssm_d_conv = 0
print_info: ssm_d_inner = 0
print_info: ssm_d_state = 0
print_info: ssm_dt_rank = 0
print_info: ssm_dt_b_c_rms = 0
print_info: model type = 12B
print_info: model params = 11.77 B
print_info: general.name = Gemma 3
print_info: vocab type = SPM
print_info: n_vocab = 262208
print_info: n_merges = 0
print_info: BOS token = 2 ''
print_info: EOS token = 106 '<end_of_turn>'
print_info: EOT token = 106 '<end_of_turn>'
print_info: UNK token = 3 ''
print_info: PAD token = 0 ''
print_info: LF token = 248 '<0x0A>'
print_info: EOG token = 106 '<end_of_turn>'
print_info: max token length = 48
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 48 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 49/49 layers to GPU
load_tensors: CUDA0 model buffer size = 8047.63 MiB
load_tensors: CPU_Mapped model buffer size = 787.69 MiB
.....................................................................................
llama_init_from_model: n_seq_max = 1
llama_init_from_model: n_ctx = 4096
llama_init_from_model: n_ctx_per_seq = 4096
llama_init_from_model: n_batch = 1024
llama_init_from_model: n_ubatch = 512
llama_init_from_model: flash_attn = 1
llama_init_from_model: freq_base = 1000000.0
llama_init_from_model: freq_scale = 0.125
llama_init_from_model: n_ctx_per_seq (4096) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
llama_kv_cache_init: kv_size = 4096, offload = 1, type_k = 'q8_0', type_v = 'q8_0', n_layer = 48, can_shift = 1
llama_kv_cache_init: CUDA0 KV buffer size = 816.00 MiB
llama_init_from_model: KV self size = 816.00 MiB, K (q8_0): 408.00 MiB, V (q8_0): 408.00 MiB
llama_init_from_model: CUDA_Host output buffer size = 1.00 MiB
llama_init_from_model: CUDA0 compute buffer size = 519.62 MiB
llama_init_from_model: CUDA_Host compute buffer size = 45.01 MiB
llama_init_from_model: graph nodes = 1737
llama_init_from_model: graph splits = 98
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv init: initializing slots, n_slots = 1
slot init: id 0 | task -1 | new slot n_ctx_slot = 4096
main: model loaded
main: chat template, chat_template: {{ bos_token }}
{%- if messages[0]['role'] == 'system' -%}
{%- if messages[0]['content'] is string -%}
{%- set first_user_prefix = messages[0]['content'] + '

' -%}
{%- else -%}
{%- set first_user_prefix = messages[0]['content'][0]['text'] + '

' -%}
{%- endif -%}
{%- set loop_messages = messages[1:] -%}
{%- else -%}
{%- set first_user_prefix = "" -%}
{%- set loop_messages = messages -%}
{%- endif -%}
{%- for message in loop_messages -%}
{%- if (message['role'] == 'user') != (loop.index0 % 2 == 0) -%}
{{ raise_exception("Conversation roles must alternate user/assistant/user/assistant/...") }}
{%- endif -%}
{%- if (message['role'] == 'assistant') -%}
{%- set role = "model" -%}
{%- else -%}
{%- set role = message['role'] -%}
{%- endif -%}
{{ '<start_of_turn>' + role + '
' + (first_user_prefix if loop.first else "") }}
{%- if message['content'] is string -%}
{{ message['content'] | trim }}
{%- elif message['content'] is iterable -%}
{%- for item in message['content'] -%}
{%- if item['type'] == 'image' -%}
{{ '<start_of_image>' }}
{%- elif item['type'] == 'text' -%}
{{ item['text'] | trim }}
{%- endif -%}
{%- endfor -%}
{%- else -%}
{{ raise_exception("Invalid content type") }}
{%- endif -%}
{{ '<end_of_turn>
' }}
{%- endfor -%}
{%- if add_generation_prompt -%}
{{'<start_of_turn>model
'}}
{%- endif -%}
, example_format: '<start_of_turn>user
You are a helpful assistant

Hello<end_of_turn>
<start_of_turn>model
Hi there<end_of_turn>
<start_of_turn>user
How are you?<end_of_turn>
<start_of_turn>model
'
main: server is listening on http://127.0.0.1:7777 - starting the main loop
srv update_slots: all slots are idle
srv log_server_r: request: GET / 127.0.0.1 200
srv log_server_r: request: GET /favicon.ico 127.0.0.1 404
srv params_from_: Chat format: Content-only
slot launch_slot_: id 0 | task 0 | processing task
slot update_slots: id 0 | task 0 | new prompt, n_ctx_slot = 4096, n_keep = 0, n_prompt_tokens = 2505
slot update_slots: id 0 | task 0 | kv cache rm [0, end)
slot update_slots: id 0 | task 0 | prompt processing progress, n_past = 1024, n_tokens = 1024, progress = 0.408782
slot update_slots: id 0 | task 0 | kv cache rm [1024, end)
slot update_slots: id 0 | task 0 | prompt processing progress, n_past = 2048, n_tokens = 1024, progress = 0.817565
slot update_slots: id 0 | task 0 | kv cache rm [2048, end)
slot update_slots: id 0 | task 0 | prompt processing progress, n_past = 2505, n_tokens = 457, progress = 1.000000
slot update_slots: id 0 | task 0 | prompt done, n_past = 2505, n_tokens = 457
slot release: id 0 | task 0 | stop processing: n_past = 3225, truncated = 0
slot print_timing: id 0 | task 0 |
prompt eval time = 22698.14 ms / 2505 tokens ( 9.06 ms per token, 110.36 tokens per second)
eval time = 78316.51 ms / 721 tokens ( 108.62 ms per token, 9.21 tokens per second)
total time = 101014.65 ms / 3226 tokens
srv update_slots: all slots are idle
srv log_server_r: request: POST /v1/chat/completions 127.0.0.1 200

slaren · 2025-03-12T13:10:08Z

llama_init_from_model: graph splits = 98

The large number of graph splits indicates that there is some operation that is not supported by the CUDA backend, and is being run on the CPU. If you set the environment variable GGML_SCHED_DEBUG=2 and run with -v, you should get a report that will show which operations are being run on the CPU.

Bearsaerker · 2025-03-12T13:26:52Z

This is the log with your env variables and -v

SPLIT #0: CPU # 0 inputs

node # 0 ( GET_ROWS): inp_embd ( 15K) [ CPU ]: token_embd.weight ( 787M) [ CPU ] inp_tokens ( 0K) [ CPU ]

SPLIT #1: CUDA0 # 3 inputs: [inp_embd ( 15K)] [inp_pos ( 0K)] [KQ_mask_swa ( 64K)]

node # 1 ( SCALE): inp_scaled ( 15K) [CUDA0 ]: CUDA0#inp_embd#0 ( 15K) [ NULL ]
node # 2 ( RMS_NORM): norm-0 ( 15K) [CUDA0 ]: inp_scaled ( 15K) [CUDA0 ]
node # 3 ( MUL): attn_norm-0 ( 15K) [CUDA0 ]: norm-0 ( 15K) [CUDA0 ] blk.0.attn_norm.weig ( 15K) [CUDA0 ]
node # 4 ( MUL_MAT): Qcur-0 ( 16K) [CUDA0 ]: blk.0.attn_q.weight ( 8M) [CUDA0 ] attn_norm-0 ( 15K) [CUDA0 ]
node # 6 ( RMS_NORM): norm-0 ( 16K) [CUDA0 ]: Qcur-0 (reshaped) ( 16K) [CUDA0 ]
node # 7 ( MUL): Qcur_normed-0 ( 16K) [CUDA0 ]: norm-0 ( 16K) [CUDA0 ] blk.0.attn_q_norm.we ( 1K) [CUDA0 ]
node # 8 ( ROPE): Qcur-0 ( 16K) [CUDA0 ]: Qcur_normed-0 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node # 9 ( MUL_MAT): Kcur-0 ( 8K) [CUDA0 ]: blk.0.attn_k.weight ( 4M) [CUDA0 ] attn_norm-0 ( 15K) [CUDA0 ]
node # 11 ( RMS_NORM): norm-0 ( 8K) [CUDA0 ]: Kcur-0 (reshaped) ( 8K) [CUDA0 ]
node # 12 ( MUL): Kcur_normed-0 ( 8K) [CUDA0 ]: norm-0 ( 8K) [CUDA0 ] blk.0.attn_k_norm.we ( 1K) [CUDA0 ]
node # 13 ( ROPE): Kcur-0 ( 8K) [CUDA0 ]: Kcur_normed-0 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node # 14 ( MUL_MAT): Vcur-0 ( 8K) [CUDA0 ]: blk.0.attn_v.weight ( 6M) [CUDA0 ] attn_norm-0 ( 15K) [CUDA0 ]
node # 16 ( CPY): k_cache_view-0 (copy ( 2K) [CUDA0 ]: Kcur-0 ( 8K) [CUDA0 ] k_cache_view-0 ( 2K) [CUDA0 ]
node # 18 ( CPY): v_cache_view-0 (copy ( 2K) [CUDA0 ]: Vcur-0 ( 8K) [CUDA0 ] v_cache_view-0 ( 2K) [CUDA0 ]
node # 22 ( CPY): KQ_mask_swa (copy) ( 32K) [CUDA0 ]: CUDA0#KQ_mask_swa#0 ( 64K) [ NULL ] KQ_mask_swa (copy) ( 32K) [CUDA0 ]

SPLIT #2: CPU # 4 inputs: [q-0 ( 16K)] [k-0 ( 544K)] [v-0 ( 544K)] [KQ_mask_swa (copy) ( 32K)]

node # 23 (FLASH_ATTN): node_23 ( 16K) [ CPU ]: CPU#q-0#0 ( 16K) [ NULL ] CPU#k-0#0 ( 544K) [ NULL ] CPU#v-0#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #3: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node # 25 ( MUL_MAT): node # 26 ( RMS_NORM): node # 27 ( node # 28 ( ADD): node # 29 ( RMS_NORM): node # 30 ( MUL): node # 31 ( MUL_MAT): node # 32 ( UNARY): node # 33 ( MUL_MAT): node # 34 ( MUL): node # 35 ( MUL_MAT): node # 36 ( RMS_NORM): node # 37 ( MUL): node # 38 ( ADD): node # 39 ( RMS_NORM): node # 40 ( MUL): node # 41 ( MUL_MAT): node # 43 ( RMS_NORM): node # 44 ( MUL): node # 45 ( ROPE): node # 46 ( MUL_MAT): node # 48 ( RMS_NORM): node # 49 ( MUL): node # 50 ( ROPE): node # 51 ( MUL_MAT): node # 53 ( node # 55 ( kqv_out-0 ( 15K) [CUDA0 ]: blk.0.attn_output.we ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
norm-0 ( 15K) [CUDA0 ]: kqv_out-0 ( 15K) [CUDA0 ]
MUL): attn_post_norm-0 ( 15K) [CUDA0 ]: norm-0 ( 15K) [CUDA0 ] blk.0.post_attention ( 15K) [CUDA0 ]
sa_out-0 ( 15K) [CUDA0 ]: attn_post_norm-0 ( 15K) [CUDA0 ] inp_scaled ( 15K) [CUDA0 ]
norm-0 ( 15K) [CUDA0 ]: sa_out-0 ( 15K) [CUDA0 ]
ffn_norm-0 ( 15K) [CUDA0 ]: norm-0 ( 15K) [CUDA0 ] blk.0.ffn_norm.weigh ( 15K) [CUDA0 ]
ffn_gate-0 ( 60K) [CUDA0 ]: blk.0.ffn_gate.weigh ( 31M) [CUDA0 ] ffn_norm-0 ( 15K) [CUDA0 ]
ffn_gelu-0 ( 60K) [CUDA0 ]: ffn_gate-0 ( 60K) [CUDA0 ]
ffn_up-0 ( 60K) [CUDA0 ]: blk.0.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-0 ( 15K) [CUDA0 ]
ffn_gate_par-0 ( 60K) [CUDA0 ]: ffn_gelu-0 ( 60K) [CUDA0 ] ffn_up-0 ( 60K) [CUDA0 ]
ffn_out-0 ( 15K) [CUDA0 ]: blk.0.ffn_down.weigh ( 46M) [CUDA0 ] ffn_gate_par-0 ( 60K) [CUDA0 ]
norm ( 15K) [CUDA0 ]: ffn_out-0 ( 15K) [CUDA0 ]
ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.0.post_ffw_norm. ( 15K) [CUDA0 ]
l_out-0 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-0 ( 15K) [CUDA0 ]
norm-1 ( 15K) [CUDA0 ]: l_out-0 ( 15K) [CUDA0 ]
attn_norm-1 ( 15K) [CUDA0 ]: norm-1 ( 15K) [CUDA0 ] blk.1.attn_norm.weig ( 15K) [CUDA0 ]
Qcur-1 ( 16K) [CUDA0 ]: blk.1.attn_q.weight ( 8M) [CUDA0 ] attn_norm-1 ( 15K) [CUDA0 ]
norm-1 ( 16K) [CUDA0 ]: Qcur-1 (reshaped) ( 16K) [CUDA0 ]
Qcur_normed-1 ( 16K) [CUDA0 ]: norm-1 ( 16K) [CUDA0 ] blk.1.attn_q_norm.we ( 1K) [CUDA0 ]
Qcur-1 ( 16K) [CUDA0 ]: Qcur_normed-1 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
Kcur-1 ( 8K) [CUDA0 ]: blk.1.attn_k.weight ( 4M) [CUDA0 ] attn_norm-1 ( 15K) [CUDA0 ]
norm-1 ( 8K) [CUDA0 ]: Kcur-1 (reshaped) ( 8K) [CUDA0 ]
Kcur_normed-1 ( 8K) [CUDA0 ]: norm-1 ( 8K) [CUDA0 ] blk.1.attn_k_norm.we ( 1K) [CUDA0 ]
Kcur-1 ( 8K) [CUDA0 ]: Kcur_normed-1 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
Vcur-1 ( 8K) [CUDA0 ]: blk.1.attn_v.weight ( 6M) [CUDA0 ] attn_norm-1 ( 15K) [CUDA0 ]
CPY): k_cache_view-1 (copy ( 2K) [CUDA0 ]: Kcur-1 ( 8K) [CUDA0 ] k_cache_view-1 ( 2K) [CUDA0 ]
CPY): v_cache_view-1 (copy ( 2K) [CUDA0 ]: Vcur-1 ( 8K) [CUDA0 ] v_cache_view-1 ( 2K) [CUDA0 ]

SPLIT #4: CPU # 3 inputs: [q-1 ( 16K)] [k-1 ( 544K)] [v-1 ( 544K)]

node # 59 (FLASH_ATTN): node_59 ( 16K) [ CPU ]: CPU#q-1#0 ( 16K) [ NULL ] CPU#k-1#0 ( 544K) [ NULL ] CPU#v-1#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #5: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node # 61 ( MUL_MAT): kqv_out-1 ( 15K) [CUDA0 ]: blk.1.attn_output.we ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node # 62 ( RMS_NORM): norm-1 ( 15K) [CUDA0 ]: kqv_out-1 ( 15K) [CUDA0 ]
node # 63 ( MUL): attn_post_norm-1 ( 15K) [CUDA0 ]: norm-1 ( 15K) [CUDA0 ] blk.1.post_attention ( 15K) [CUDA0 ]
node # 64 ( ADD): sa_out-1 ( 15K) [CUDA0 ]: attn_post_norm-1 ( 15K) [CUDA0 ] l_out-0 ( 15K) [CUDA0 ]
node # 65 ( RMS_NORM): norm-1 ( 15K) [CUDA0 ]: sa_out-1 ( 15K) [CUDA0 ]
node # 66 ( MUL): ffn_norm-1 ( 15K) [CUDA0 ]: norm-1 ( 15K) [CUDA0 ] blk.1.ffn_norm.weigh ( 15K) [CUDA0 ]
node # 67 ( MUL_MAT): ffn_gate-1 ( 60K) [CUDA0 ]: blk.1.ffn_gate.weigh ( 31M) [CUDA0 ] ffn_norm-1 ( 15K) [CUDA0 ]
node # 68 ( UNARY): ffn_gelu-1 ( 60K) [CUDA0 ]: ffn_gate-1 ( 60K) [CUDA0 ]
node # 69 ( MUL_MAT): ffn_up-1 ( 60K) [CUDA0 ]: blk.1.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-1 ( 15K) [CUDA0 ]
node # 70 ( MUL): ffn_gate_par-1 ( 60K) [CUDA0 ]: ffn_gelu-1 ( 60K) [CUDA0 ] ffn_up-1 ( 60K) [CUDA0 ]
node # 71 ( MUL_MAT): ffn_out-1 ( 15K) [CUDA0 ]: blk.1.ffn_down.weigh ( 46M) [CUDA0 ] ffn_gate_par-1 ( 60K) [CUDA0 ]
node # 72 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-1 ( 15K) [CUDA0 ]
node # 73 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.1.post_ffw_norm. ( 15K) [CUDA0 ]
node # 74 ( ADD): l_out-1 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-1 ( 15K) [CUDA0 ]
node # 75 ( RMS_NORM): norm-2 ( 15K) [CUDA0 ]: l_out-1 ( 15K) [CUDA0 ]
node # 76 ( MUL): attn_norm-2 ( 15K) [CUDA0 ]: norm-2 ( 15K) [CUDA0 ] blk.2.attn_norm.weig ( 15K) [CUDA0 ]
node # 77 ( MUL_MAT): Qcur-2 ( 16K) [CUDA0 ]: blk.2.attn_q.weight ( 8M) [CUDA0 ] attn_norm-2 ( 15K) [CUDA0 ]
node # 79 ( RMS_NORM): norm-2 ( 16K) [CUDA0 ]: Qcur-2 (reshaped) ( 16K) [CUDA0 ]
node # 80 ( MUL): Qcur_normed-2 ( 16K) [CUDA0 ]: norm-2 ( 16K) [CUDA0 ] blk.2.attn_q_norm.we ( 1K) [CUDA0 ]
node # 81 ( ROPE): Qcur-2 ( 16K) [CUDA0 ]: Qcur_normed-2 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node # 82 ( MUL_MAT): Kcur-2 ( 8K) [CUDA0 ]: blk.2.attn_k.weight ( 4M) [CUDA0 ] attn_norm-2 ( 15K) [CUDA0 ]
node # 84 ( RMS_NORM): norm-2 ( 8K) [CUDA0 ]: Kcur-2 (reshaped) ( 8K) [CUDA0 ]
node # 85 ( MUL): Kcur_normed-2 ( 8K) [CUDA0 ]: norm-2 ( 8K) [CUDA0 ] blk.2.attn_k_norm.we ( 1K) [CUDA0 ]
node # 86 ( ROPE): Kcur-2 ( 8K) [CUDA0 ]: Kcur_normed-2 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node # 87 ( MUL_MAT): Vcur-2 ( 8K) [CUDA0 ]: blk.2.attn_v.weight ( 6M) [CUDA0 ] attn_norm-2 ( 15K) [CUDA0 ]
node # 89 ( CPY): k_cache_view-2 (copy ( 2K) [CUDA0 ]: Kcur-2 ( 8K) [CUDA0 ] k_cache_view-2 ( 2K) [CUDA0 ]
node # 91 ( CPY): v_cache_view-2 (copy ( 2K) [CUDA0 ]: Vcur-2 ( 8K) [CUDA0 ] v_cache_view-2 ( 2K) [CUDA0 ]

SPLIT #6: CPU # 3 inputs: [q-2 ( 16K)] [k-2 ( 544K)] [v-2 ( 544K)]

node # 95 (FLASH_ATTN): node_95 ( 16K) [ CPU ]: CPU#q-2#0 ( 16K) [ NULL ] CPU#k-2#0 ( 544K) [ NULL ] CPU#v-2#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #7: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node # 97 ( MUL_MAT): kqv_out-2 ( 15K) [CUDA0 ]: blk.2.attn_output.we ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node # 98 ( RMS_NORM): norm-2 ( 15K) [CUDA0 ]: kqv_out-2 ( 15K) [CUDA0 ]
node # 99 ( MUL): attn_post_norm-2 ( 15K) [CUDA0 ]: norm-2 ( 15K) [CUDA0 ] blk.2.post_attention ( 15K) [CUDA0 ]
node #100 ( ADD): sa_out-2 ( 15K) [CUDA0 ]: attn_post_norm-2 ( 15K) [CUDA0 ] l_out-1 ( 15K) [CUDA0 ]
node #101 ( RMS_NORM): norm-2 ( 15K) [CUDA0 ]: sa_out-2 ( 15K) [CUDA0 ]
node #102 ( MUL): ffn_norm-2 ( 15K) [CUDA0 ]: norm-2 ( 15K) [CUDA0 ] blk.2.ffn_norm.weigh ( 15K) [CUDA0 ]
node #103 ( MUL_MAT): ffn_gate-2 ( 60K) [CUDA0 ]: blk.2.ffn_gate.weigh ( 31M) [CUDA0 ] ffn_norm-2 ( 15K) [CUDA0 ]
node #104 ( UNARY): ffn_gelu-2 ( 60K) [CUDA0 ]: ffn_gate-2 ( 60K) [CUDA0 ]
node #105 ( MUL_MAT): ffn_up-2 ( 60K) [CUDA0 ]: blk.2.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-2 ( 15K) [CUDA0 ]
node #106 ( MUL): ffn_gate_par-2 ( 60K) [CUDA0 ]: ffn_gelu-2 ( 60K) [CUDA0 ] ffn_up-2 ( 60K) [CUDA0 ]
node #107 ( MUL_MAT): ffn_out-2 ( 15K) [CUDA0 ]: blk.2.ffn_down.weigh ( 46M) [CUDA0 ] ffn_gate_par-2 ( 60K) [CUDA0 ]
node #108 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-2 ( 15K) [CUDA0 ]
node #109 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.2.post_ffw_norm. ( 15K) [CUDA0 ]
node #110 ( ADD): l_out-2 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-2 ( 15K) [CUDA0 ]
node #111 ( RMS_NORM): norm-3 ( 15K) [CUDA0 ]: l_out-2 ( 15K) [CUDA0 ]
node #112 ( MUL): attn_norm-3 ( 15K) [CUDA0 ]: norm-3 ( 15K) [CUDA0 ] blk.3.attn_norm.weig ( 15K) [CUDA0 ]
node #113 ( MUL_MAT): Qcur-3 ( 16K) [CUDA0 ]: blk.3.attn_q.weight ( 8M) [CUDA0 ] attn_norm-3 ( 15K) [CUDA0 ]
node #115 ( RMS_NORM): norm-3 ( 16K) [CUDA0 ]: Qcur-3 (reshaped) ( 16K) [CUDA0 ]
node #116 ( MUL): Qcur_normed-3 ( 16K) [CUDA0 ]: norm-3 ( 16K) [CUDA0 ] blk.3.attn_q_norm.we ( 1K) [CUDA0 ]
node #117 ( ROPE): Qcur-3 ( 16K) [CUDA0 ]: Qcur_normed-3 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #118 ( MUL_MAT): Kcur-3 ( 8K) [CUDA0 ]: blk.3.attn_k.weight ( 4M) [CUDA0 ] attn_norm-3 ( 15K) [CUDA0 ]
node #120 ( RMS_NORM): norm-3 ( 8K) [CUDA0 ]: Kcur-3 (reshaped) ( 8K) [CUDA0 ]
node #121 ( MUL): Kcur_normed-3 ( 8K) [CUDA0 ]: norm-3 ( 8K) [CUDA0 ] blk.3.attn_k_norm.we ( 1K) [CUDA0 ]
node #122 ( ROPE): Kcur-3 ( 8K) [CUDA0 ]: Kcur_normed-3 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #123 ( MUL_MAT): Vcur-3 ( 8K) [CUDA0 ]: blk.3.attn_v.weight ( 6M) [CUDA0 ] attn_norm-3 ( 15K) [CUDA0 ]
node #125 ( CPY): k_cache_view-3 (copy ( 2K) [CUDA0 ]: Kcur-3 ( 8K) [CUDA0 ] k_cache_view-3 ( 2K) [CUDA0 ]
node #127 ( CPY): v_cache_view-3 (copy ( 2K) [CUDA0 ]: Vcur-3 ( 8K) [CUDA0 ] v_cache_view-3 ( 2K) [CUDA0 ]

SPLIT #8: CPU # 3 inputs: [q-3 ( 16K)] [k-3 ( 544K)] [v-3 ( 544K)]

node #131 (FLASH_ATTN): node_131 ( 16K) [ CPU ]: CPU#q-3#0 ( 16K) [ NULL ] CPU#k-3#0 ( 544K) [ NULL ] CPU#v-3#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #9: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #133 ( MUL_MAT): kqv_out-3 ( 15K) [CUDA0 ]: blk.3.attn_output.we ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #134 ( RMS_NORM): norm-3 ( 15K) [CUDA0 ]: kqv_out-3 ( 15K) [CUDA0 ]
node #135 ( MUL): attn_post_norm-3 ( 15K) [CUDA0 ]: norm-3 ( 15K) [CUDA0 ] blk.3.post_attention ( 15K) [CUDA0 ]
node #136 ( ADD): sa_out-3 ( 15K) [CUDA0 ]: attn_post_norm-3 ( 15K) [CUDA0 ] l_out-2 ( 15K) [CUDA0 ]
node #137 ( RMS_NORM): norm-3 ( 15K) [CUDA0 ]: sa_out-3 ( 15K) [CUDA0 ]
node #138 ( MUL): ffn_norm-3 ( 15K) [CUDA0 ]: norm-3 ( 15K) [CUDA0 ] blk.3.ffn_norm.weigh ( 15K) [CUDA0 ]
node #139 ( MUL_MAT): ffn_gate-3 ( 60K) [CUDA0 ]: blk.3.ffn_gate.weigh ( 31M) [CUDA0 ] ffn_norm-3 ( 15K) [CUDA0 ]
node #140 ( UNARY): ffn_gelu-3 ( 60K) [CUDA0 ]: ffn_gate-3 ( 60K) [CUDA0 ]
node #141 ( MUL_MAT): ffn_up-3 ( 60K) [CUDA0 ]: blk.3.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-3 ( 15K) [CUDA0 ]
node #142 ( MUL): ffn_gate_par-3 ( 60K) [CUDA0 ]: ffn_gelu-3 ( 60K) [CUDA0 ] ffn_up-3 ( 60K) [CUDA0 ]
node #143 ( MUL_MAT): ffn_out-3 ( 15K) [CUDA0 ]: blk.3.ffn_down.weigh ( 46M) [CUDA0 ] ffn_gate_par-3 ( 60K) [CUDA0 ]
node #144 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-3 ( 15K) [CUDA0 ]
node #145 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.3.post_ffw_norm. ( 15K) [CUDA0 ]
node #146 ( ADD): l_out-3 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-3 ( 15K) [CUDA0 ]
node #147 ( RMS_NORM): norm-4 ( 15K) [CUDA0 ]: l_out-3 ( 15K) [CUDA0 ]
node #148 ( MUL): attn_norm-4 ( 15K) [CUDA0 ]: norm-4 ( 15K) [CUDA0 ] blk.4.attn_norm.weig ( 15K) [CUDA0 ]
node #149 ( MUL_MAT): Qcur-4 ( 16K) [CUDA0 ]: blk.4.attn_q.weight ( 8M) [CUDA0 ] attn_norm-4 ( 15K) [CUDA0 ]
node #151 ( RMS_NORM): norm-4 ( 16K) [CUDA0 ]: Qcur-4 (reshaped) ( 16K) [CUDA0 ]
node #152 ( MUL): Qcur_normed-4 ( 16K) [CUDA0 ]: norm-4 ( 16K) [CUDA0 ] blk.4.attn_q_norm.we ( 1K) [CUDA0 ]
node #153 ( ROPE): Qcur-4 ( 16K) [CUDA0 ]: Qcur_normed-4 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #154 ( MUL_MAT): Kcur-4 ( 8K) [CUDA0 ]: blk.4.attn_k.weight ( 4M) [CUDA0 ] attn_norm-4 ( 15K) [CUDA0 ]
node #156 ( RMS_NORM): norm-4 ( 8K) [CUDA0 ]: Kcur-4 (reshaped) ( 8K) [CUDA0 ]
node #157 ( MUL): Kcur_normed-4 ( 8K) [CUDA0 ]: norm-4 ( 8K) [CUDA0 ] blk.4.attn_k_norm.we ( 1K) [CUDA0 ]
node #158 ( ROPE): Kcur-4 ( 8K) [CUDA0 ]: Kcur_normed-4 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #159 ( MUL_MAT): Vcur-4 ( 8K) [CUDA0 ]: blk.4.attn_v.weight ( 6M) [CUDA0 ] attn_norm-4 ( 15K) [CUDA0 ]
node #161 ( CPY): k_cache_view-4 (copy ( 2K) [CUDA0 ]: Kcur-4 ( 8K) [CUDA0 ] k_cache_view-4 ( 2K) [CUDA0 ]
node #163 ( CPY): v_cache_view-4 (copy ( 2K) [CUDA0 ]: Vcur-4 ( 8K) [CUDA0 ] v_cache_view-4 ( 2K) [CUDA0 ]

SPLIT #10: CPU # 3 inputs: [q-4 ( 16K)] [k-4 ( 544K)] [v-4 ( 544K)]

node #167 (FLASH_ATTN): node_167 ( 16K) [ CPU ]: CPU#q-4#0 ( 16K) [ NULL ] CPU#k-4#0 ( 544K) [ NULL ] CPU#v-4#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #11: CUDA0 # 2 inputs: [ (reshaped) ( 16K)] [KQ_mask ( 64K)]

node #169 ( MUL_MAT): kqv_out-4 ( 15K) [CUDA0 ]: blk.4.attn_output.we ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #170 ( RMS_NORM): norm-4 ( 15K) [CUDA0 ]: kqv_out-4 ( 15K) [CUDA0 ]
node #171 ( MUL): attn_post_norm-4 ( 15K) [CUDA0 ]: norm-4 ( 15K) [CUDA0 ] blk.4.post_attention ( 15K) [CUDA0 ]
node #172 ( ADD): sa_out-4 ( 15K) [CUDA0 ]: attn_post_norm-4 ( 15K) [CUDA0 ] l_out-3 ( 15K) [CUDA0 ]
node #173 ( RMS_NORM): norm-4 ( 15K) [CUDA0 ]: sa_out-4 ( 15K) [CUDA0 ]
node #174 ( MUL): ffn_norm-4 ( 15K) [CUDA0 ]: norm-4 ( 15K) [CUDA0 ] blk.4.ffn_norm.weigh ( 15K) [CUDA0 ]
node #175 ( MUL_MAT): ffn_gate-4 ( 60K) [CUDA0 ]: blk.4.ffn_gate.weigh ( 31M) [CUDA0 ] ffn_norm-4 ( 15K) [CUDA0 ]
node #176 ( UNARY): ffn_gelu-4 ( 60K) [CUDA0 ]: ffn_gate-4 ( 60K) [CUDA0 ]
node #177 ( MUL_MAT): ffn_up-4 ( 60K) [CUDA0 ]: blk.4.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-4 ( 15K) [CUDA0 ]
node #178 ( MUL): ffn_gate_par-4 ( 60K) [CUDA0 ]: ffn_gelu-4 ( 60K) [CUDA0 ] ffn_up-4 ( 60K) [CUDA0 ]
node #179 ( MUL_MAT): ffn_out-4 ( 15K) [CUDA0 ]: blk.4.ffn_down.weigh ( 46M) [CUDA0 ] ffn_gate_par-4 ( 60K) [CUDA0 ]
node #180 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-4 ( 15K) [CUDA0 ]
node #181 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.4.post_ffw_norm. ( 15K) [CUDA0 ]
node #182 ( ADD): l_out-4 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-4 ( 15K) [CUDA0 ]
node #183 ( RMS_NORM): norm-5 ( 15K) [CUDA0 ]: l_out-4 ( 15K) [CUDA0 ]
node #184 ( MUL): attn_norm-5 ( 15K) [CUDA0 ]: norm-5 ( 15K) [CUDA0 ] blk.5.attn_norm.weig ( 15K) [CUDA0 ]
node #185 ( MUL_MAT): Qcur-5 ( 16K) [CUDA0 ]: blk.5.attn_q.weight ( 8M) [CUDA0 ] attn_norm-5 ( 15K) [CUDA0 ]
node #187 ( RMS_NORM): norm-5 ( 16K) [CUDA0 ]: Qcur-5 (reshaped) ( 16K) [CUDA0 ]
node #188 ( MUL): Qcur_normed-5 ( 16K) [CUDA0 ]: norm-5 ( 16K) [CUDA0 ] blk.5.attn_q_norm.we ( 1K) [CUDA0 ]
node #189 ( ROPE): Qcur-5 ( 16K) [CUDA0 ]: Qcur_normed-5 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #190 ( MUL_MAT): Kcur-5 ( 8K) [CUDA0 ]: blk.5.attn_k.weight ( 4M) [CUDA0 ] attn_norm-5 ( 15K) [CUDA0 ]
node #192 ( RMS_NORM): norm-5 ( 8K) [CUDA0 ]: Kcur-5 (reshaped) ( 8K) [CUDA0 ]
node #193 ( MUL): Kcur_normed-5 ( 8K) [CUDA0 ]: norm-5 ( 8K) [CUDA0 ] blk.5.attn_k_norm.we ( 1K) [CUDA0 ]
node #194 ( ROPE): Kcur-5 ( 8K) [CUDA0 ]: Kcur_normed-5 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #195 ( MUL_MAT): Vcur-5 ( 8K) [CUDA0 ]: blk.5.attn_v.weight ( 6M) [CUDA0 ] attn_norm-5 ( 15K) [CUDA0 ]
node #197 ( CPY): k_cache_view-5 (copy ( 2K) [CUDA0 ]: Kcur-5 ( 8K) [CUDA0 ] k_cache_view-5 ( 2K) [CUDA0 ]
node #199 ( CPY): v_cache_view-5 (copy ( 2K) [CUDA0 ]: Vcur-5 ( 8K) [CUDA0 ] v_cache_view-5 ( 2K) [CUDA0 ]
node #203 ( CPY): KQ_mask (copy) ( 32K) [CUDA0 ]: CUDA0#KQ_mask#0 ( 64K) [ NULL ] KQ_mask (copy) ( 32K) [CUDA0 ]

SPLIT #12: CPU # 4 inputs: [q-5 ( 16K)] [k-5 ( 544K)] [v-5 ( 544K)] [KQ_mask (copy) ( 32K)]

node #204 (FLASH_ATTN): node_204 ( 16K) [ CPU ]: CPU#q-5#0 ( 16K) [ NULL ] CPU#k-5#0 ( 544K) [ NULL ] CPU#v-5#0 ( 544K) [ NULL ] CPU#KQ_mask (copy)#0 ( 32K) [ NULL ]

SPLIT #13: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #206 ( MUL_MAT): kqv_out-5 ( 15K) [CUDA0 ]: blk.5.attn_output.we ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #207 ( RMS_NORM): norm-5 ( 15K) [CUDA0 ]: kqv_out-5 ( 15K) [CUDA0 ]
node #208 ( MUL): attn_post_norm-5 ( 15K) [CUDA0 ]: norm-5 ( 15K) [CUDA0 ] blk.5.post_attention ( 15K) [CUDA0 ]
node #209 ( ADD): sa_out-5 ( 15K) [CUDA0 ]: attn_post_norm-5 ( 15K) [CUDA0 ] l_out-4 ( 15K) [CUDA0 ]
node #210 ( RMS_NORM): norm-5 ( 15K) [CUDA0 ]: sa_out-5 ( 15K) [CUDA0 ]
node #211 ( MUL): ffn_norm-5 ( 15K) [CUDA0 ]: norm-5 ( 15K) [CUDA0 ] blk.5.ffn_norm.weigh ( 15K) [CUDA0 ]
node #212 ( MUL_MAT): ffn_gate-5 ( 60K) [CUDA0 ]: blk.5.ffn_gate.weigh ( 31M) [CUDA0 ] ffn_norm-5 ( 15K) [CUDA0 ]
node #213 ( UNARY): ffn_gelu-5 ( 60K) [CUDA0 ]: ffn_gate-5 ( 60K) [CUDA0 ]
node #214 ( MUL_MAT): ffn_up-5 ( 60K) [CUDA0 ]: blk.5.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-5 ( 15K) [CUDA0 ]
node #215 ( MUL): ffn_gate_par-5 ( 60K) [CUDA0 ]: ffn_gelu-5 ( 60K) [CUDA0 ] ffn_up-5 ( 60K) [CUDA0 ]
node #216 ( MUL_MAT): ffn_out-5 ( 15K) [CUDA0 ]: blk.5.ffn_down.weigh ( 46M) [CUDA0 ] ffn_gate_par-5 ( 60K) [CUDA0 ]
node #217 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-5 ( 15K) [CUDA0 ]
node #218 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.5.post_ffw_norm. ( 15K) [CUDA0 ]
node #219 ( ADD): l_out-5 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-5 ( 15K) [CUDA0 ]
node #220 ( RMS_NORM): norm-6 ( 15K) [CUDA0 ]: l_out-5 ( 15K) [CUDA0 ]
node #221 ( MUL): attn_norm-6 ( 15K) [CUDA0 ]: norm-6 ( 15K) [CUDA0 ] blk.6.attn_norm.weig ( 15K) [CUDA0 ]
node #222 ( MUL_MAT): Qcur-6 ( 16K) [CUDA0 ]: blk.6.attn_q.weight ( 8M) [CUDA0 ] attn_norm-6 ( 15K) [CUDA0 ]
node #224 ( RMS_NORM): norm-6 ( 16K) [CUDA0 ]: Qcur-6 (reshaped) ( 16K) [CUDA0 ]
node #225 ( MUL): Qcur_normed-6 ( 16K) [CUDA0 ]: norm-6 ( 16K) [CUDA0 ] blk.6.attn_q_norm.we ( 1K) [CUDA0 ]
node #226 ( ROPE): Qcur-6 ( 16K) [CUDA0 ]: Qcur_normed-6 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #227 ( MUL_MAT): Kcur-6 ( 8K) [CUDA0 ]: blk.6.attn_k.weight ( 4M) [CUDA0 ] attn_norm-6 ( 15K) [CUDA0 ]
node #229 ( RMS_NORM): norm-6 ( 8K) [CUDA0 ]: Kcur-6 (reshaped) ( 8K) [CUDA0 ]
node #230 ( MUL): Kcur_normed-6 ( 8K) [CUDA0 ]: norm-6 ( 8K) [CUDA0 ] blk.6.attn_k_norm.we ( 1K) [CUDA0 ]
node #231 ( ROPE): Kcur-6 ( 8K) [CUDA0 ]: Kcur_normed-6 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #232 ( MUL_MAT): Vcur-6 ( 8K) [CUDA0 ]: blk.6.attn_v.weight ( 4M) [CUDA0 ] attn_norm-6 ( 15K) [CUDA0 ]
node #234 ( CPY): k_cache_view-6 (copy ( 2K) [CUDA0 ]: Kcur-6 ( 8K) [CUDA0 ] k_cache_view-6 ( 2K) [CUDA0 ]
node #236 ( CPY): v_cache_view-6 (copy ( 2K) [CUDA0 ]: Vcur-6 ( 8K) [CUDA0 ] v_cache_view-6 ( 2K) [CUDA0 ]

SPLIT #14: CPU # 3 inputs: [q-6 ( 16K)] [k-6 ( 544K)] [v-6 ( 544K)]

node #240 (FLASH_ATTN): node_240 ( 16K) [ CPU ]: CPU#q-6#0 ( 16K) [ NULL ] CPU#k-6#0 ( 544K) [ NULL ] CPU#v-6#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #15: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #242 ( MUL_MAT): kqv_out-6 ( 15K) [CUDA0 ]: blk.6.attn_output.we ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #243 ( RMS_NORM): norm-6 ( 15K) [CUDA0 ]: kqv_out-6 ( 15K) [CUDA0 ]
node #244 ( MUL): attn_post_norm-6 ( 15K) [CUDA0 ]: norm-6 ( 15K) [CUDA0 ] blk.6.post_attention ( 15K) [CUDA0 ]
node #245 ( ADD): sa_out-6 ( 15K) [CUDA0 ]: attn_post_norm-6 ( 15K) [CUDA0 ] l_out-5 ( 15K) [CUDA0 ]
node #246 ( RMS_NORM): norm-6 ( 15K) [CUDA0 ]: sa_out-6 ( 15K) [CUDA0 ]
node #247 ( MUL): ffn_norm-6 ( 15K) [CUDA0 ]: norm-6 ( 15K) [CUDA0 ] blk.6.ffn_norm.weigh ( 15K) [CUDA0 ]
node #248 ( MUL_MAT): ffn_gate-6 ( 60K) [CUDA0 ]: blk.6.ffn_gate.weigh ( 31M) [CUDA0 ] ffn_norm-6 ( 15K) [CUDA0 ]
node #249 ( UNARY): ffn_gelu-6 ( 60K) [CUDA0 ]: ffn_gate-6 ( 60K) [CUDA0 ]
node #250 ( MUL_MAT): ffn_up-6 ( 60K) [CUDA0 ]: blk.6.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-6 ( 15K) [CUDA0 ]
node #251 ( MUL): ffn_gate_par-6 ( 60K) [CUDA0 ]: ffn_gelu-6 ( 60K) [CUDA0 ] ffn_up-6 ( 60K) [CUDA0 ]
node #252 ( MUL_MAT): ffn_out-6 ( 15K) [CUDA0 ]: blk.6.ffn_down.weigh ( 31M) [CUDA0 ] ffn_gate_par-6 ( 60K) [CUDA0 ]
node #253 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-6 ( 15K) [CUDA0 ]
node #254 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.6.post_ffw_norm. ( 15K) [CUDA0 ]
node #255 ( ADD): l_out-6 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-6 ( 15K) [CUDA0 ]
node #256 ( RMS_NORM): norm-7 ( 15K) [CUDA0 ]: l_out-6 ( 15K) [CUDA0 ]
node #257 ( MUL): attn_norm-7 ( 15K) [CUDA0 ]: norm-7 ( 15K) [CUDA0 ] blk.7.attn_norm.weig ( 15K) [CUDA0 ]
node #258 ( MUL_MAT): Qcur-7 ( 16K) [CUDA0 ]: blk.7.attn_q.weight ( 8M) [CUDA0 ] attn_norm-7 ( 15K) [CUDA0 ]
node #260 ( RMS_NORM): norm-7 ( 16K) [CUDA0 ]: Qcur-7 (reshaped) ( 16K) [CUDA0 ]
node #261 ( MUL): Qcur_normed-7 ( 16K) [CUDA0 ]: norm-7 ( 16K) [CUDA0 ] blk.7.attn_q_norm.we ( 1K) [CUDA0 ]
node #262 ( ROPE): Qcur-7 ( 16K) [CUDA0 ]: Qcur_normed-7 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #263 ( MUL_MAT): Kcur-7 ( 8K) [CUDA0 ]: blk.7.attn_k.weight ( 4M) [CUDA0 ] attn_norm-7 ( 15K) [CUDA0 ]
node #265 ( RMS_NORM): norm-7 ( 8K) [CUDA0 ]: Kcur-7 (reshaped) ( 8K) [CUDA0 ]
node #266 ( MUL): Kcur_normed-7 ( 8K) [CUDA0 ]: norm-7 ( 8K) [CUDA0 ] blk.7.attn_k_norm.we ( 1K) [CUDA0 ]
node #267 ( ROPE): Kcur-7 ( 8K) [CUDA0 ]: Kcur_normed-7 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #268 ( MUL_MAT): Vcur-7 ( 8K) [CUDA0 ]: blk.7.attn_v.weight ( 4M) [CUDA0 ] attn_norm-7 ( 15K) [CUDA0 ]
node #270 ( CPY): k_cache_view-7 (copy ( 2K) [CUDA0 ]: Kcur-7 ( 8K) [CUDA0 ] k_cache_view-7 ( 2K) [CUDA0 ]
node #272 ( CPY): v_cache_view-7 (copy ( 2K) [CUDA0 ]: Vcur-7 ( 8K) [CUDA0 ] v_cache_view-7 ( 2K) [CUDA0 ]

SPLIT #16: CPU # 3 inputs: [q-7 ( 16K)] [k-7 ( 544K)] [v-7 ( 544K)]

node #276 (FLASH_ATTN): node_276 ( 16K) [ CPU ]: CPU#q-7#0 ( 16K) [ NULL ] CPU#k-7#0 ( 544K) [ NULL ] CPU#v-7#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #17: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #278 ( MUL_MAT): kqv_out-7 ( 15K) [CUDA0 ]: blk.7.attn_output.we ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #279 ( RMS_NORM): norm-7 ( 15K) [CUDA0 ]: kqv_out-7 ( 15K) [CUDA0 ]
node #280 ( MUL): attn_post_norm-7 ( 15K) [CUDA0 ]: norm-7 ( 15K) [CUDA0 ] blk.7.post_attention ( 15K) [CUDA0 ]
node #281 ( ADD): sa_out-7 ( 15K) [CUDA0 ]: attn_post_norm-7 ( 15K) [CUDA0 ] l_out-6 ( 15K) [CUDA0 ]
node #282 ( RMS_NORM): norm-7 ( 15K) [CUDA0 ]: sa_out-7 ( 15K) [CUDA0 ]
node #283 ( MUL): ffn_norm-7 ( 15K) [CUDA0 ]: norm-7 ( 15K) [CUDA0 ] blk.7.ffn_norm.weigh ( 15K) [CUDA0 ]
node #284 ( MUL_MAT): ffn_gate-7 ( 60K) [CUDA0 ]: blk.7.ffn_gate.weigh ( 31M) [CUDA0 ] ffn_norm-7 ( 15K) [CUDA0 ]
node #285 ( UNARY): ffn_gelu-7 ( 60K) [CUDA0 ]: ffn_gate-7 ( 60K) [CUDA0 ]
node #286 ( MUL_MAT): ffn_up-7 ( 60K) [CUDA0 ]: blk.7.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-7 ( 15K) [CUDA0 ]
node #287 ( MUL): ffn_gate_par-7 ( 60K) [CUDA0 ]: ffn_gelu-7 ( 60K) [CUDA0 ] ffn_up-7 ( 60K) [CUDA0 ]
node #288 ( MUL_MAT): ffn_out-7 ( 15K) [CUDA0 ]: blk.7.ffn_down.weigh ( 31M) [CUDA0 ] ffn_gate_par-7 ( 60K) [CUDA0 ]
node #289 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-7 ( 15K) [CUDA0 ]
node #290 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.7.post_ffw_norm. ( 15K) [CUDA0 ]
node #291 ( ADD): l_out-7 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-7 ( 15K) [CUDA0 ]
node #292 ( RMS_NORM): norm-8 ( 15K) [CUDA0 ]: l_out-7 ( 15K) [CUDA0 ]
node #293 ( MUL): attn_norm-8 ( 15K) [CUDA0 ]: norm-8 ( 15K) [CUDA0 ] blk.8.attn_norm.weig ( 15K) [CUDA0 ]
node #294 ( MUL_MAT): Qcur-8 ( 16K) [CUDA0 ]: blk.8.attn_q.weight ( 8M) [CUDA0 ] attn_norm-8 ( 15K) [CUDA0 ]
node #296 ( RMS_NORM): norm-8 ( 16K) [CUDA0 ]: Qcur-8 (reshaped) ( 16K) [CUDA0 ]
node #297 ( MUL): Qcur_normed-8 ( 16K) [CUDA0 ]: norm-8 ( 16K) [CUDA0 ] blk.8.attn_q_norm.we ( 1K) [CUDA0 ]
node #298 ( ROPE): Qcur-8 ( 16K) [CUDA0 ]: Qcur_normed-8 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #299 ( MUL_MAT): Kcur-8 ( 8K) [CUDA0 ]: blk.8.attn_k.weight ( 4M) [CUDA0 ] attn_norm-8 ( 15K) [CUDA0 ]
node #301 ( RMS_NORM): norm-8 ( 8K) [CUDA0 ]: Kcur-8 (reshaped) ( 8K) [CUDA0 ]
node #302 ( MUL): Kcur_normed-8 ( 8K) [CUDA0 ]: norm-8 ( 8K) [CUDA0 ] blk.8.attn_k_norm.we ( 1K) [CUDA0 ]
node #303 ( ROPE): Kcur-8 ( 8K) [CUDA0 ]: Kcur_normed-8 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #304 ( MUL_MAT): Vcur-8 ( 8K) [CUDA0 ]: blk.8.attn_v.weight ( 6M) [CUDA0 ] attn_norm-8 ( 15K) [CUDA0 ]
node #306 ( CPY): k_cache_view-8 (copy ( 2K) [CUDA0 ]: Kcur-8 ( 8K) [CUDA0 ] k_cache_view-8 ( 2K) [CUDA0 ]
node #308 ( CPY): v_cache_view-8 (copy ( 2K) [CUDA0 ]: Vcur-8 ( 8K) [CUDA0 ] v_cache_view-8 ( 2K) [CUDA0 ]

SPLIT #18: CPU # 3 inputs: [q-8 ( 16K)] [k-8 ( 544K)] [v-8 ( 544K)]

node #312 (FLASH_ATTN): node_312 ( 16K) [ CPU ]: CPU#q-8#0 ( 16K) [ NULL ] CPU#k-8#0 ( 544K) [ NULL ] CPU#v-8#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #19: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #314 ( MUL_MAT): kqv_out-8 ( 15K) [CUDA0 ]: blk.8.attn_output.we ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #315 ( RMS_NORM): norm-8 ( 15K) [CUDA0 ]: kqv_out-8 ( 15K) [CUDA0 ]
node #316 ( MUL): attn_post_norm-8 ( 15K) [CUDA0 ]: norm-8 ( 15K) [CUDA0 ] blk.8.post_attention ( 15K) [CUDA0 ]
node #317 ( ADD): sa_out-8 ( 15K) [CUDA0 ]: attn_post_norm-8 ( 15K) [CUDA0 ] l_out-7 ( 15K) [CUDA0 ]
node #318 ( RMS_NORM): norm-8 ( 15K) [CUDA0 ]: sa_out-8 ( 15K) [CUDA0 ]
node #319 ( MUL): ffn_norm-8 ( 15K) [CUDA0 ]: norm-8 ( 15K) [CUDA0 ] blk.8.ffn_norm.weigh ( 15K) [CUDA0 ]
node #320 ( MUL_MAT): ffn_gate-8 ( 60K) [CUDA0 ]: blk.8.ffn_gate.weigh ( 31M) [CUDA0 ] ffn_norm-8 ( 15K) [CUDA0 ]
node #321 ( UNARY): ffn_gelu-8 ( 60K) [CUDA0 ]: ffn_gate-8 ( 60K) [CUDA0 ]
node #322 ( MUL_MAT): ffn_up-8 ( 60K) [CUDA0 ]: blk.8.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-8 ( 15K) [CUDA0 ]
node #323 ( MUL): ffn_gate_par-8 ( 60K) [CUDA0 ]: ffn_gelu-8 ( 60K) [CUDA0 ] ffn_up-8 ( 60K) [CUDA0 ]
node #324 ( MUL_MAT): ffn_out-8 ( 15K) [CUDA0 ]: blk.8.ffn_down.weigh ( 46M) [CUDA0 ] ffn_gate_par-8 ( 60K) [CUDA0 ]
node #325 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-8 ( 15K) [CUDA0 ]
node #326 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.8.post_ffw_norm. ( 15K) [CUDA0 ]
node #327 ( ADD): l_out-8 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-8 ( 15K) [CUDA0 ]
node #328 ( RMS_NORM): norm-9 ( 15K) [CUDA0 ]: l_out-8 ( 15K) [CUDA0 ]
node #329 ( MUL): attn_norm-9 ( 15K) [CUDA0 ]: norm-9 ( 15K) [CUDA0 ] blk.9.attn_norm.weig ( 15K) [CUDA0 ]
node #330 ( MUL_MAT): Qcur-9 ( 16K) [CUDA0 ]: blk.9.attn_q.weight ( 8M) [CUDA0 ] attn_norm-9 ( 15K) [CUDA0 ]
node #332 ( RMS_NORM): norm-9 ( 16K) [CUDA0 ]: Qcur-9 (reshaped) ( 16K) [CUDA0 ]
node #333 ( MUL): Qcur_normed-9 ( 16K) [CUDA0 ]: norm-9 ( 16K) [CUDA0 ] blk.9.attn_q_norm.we ( 1K) [CUDA0 ]
node #334 ( ROPE): Qcur-9 ( 16K) [CUDA0 ]: Qcur_normed-9 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #335 ( MUL_MAT): Kcur-9 ( 8K) [CUDA0 ]: blk.9.attn_k.weight ( 4M) [CUDA0 ] attn_norm-9 ( 15K) [CUDA0 ]
node #337 ( RMS_NORM): norm-9 ( 8K) [CUDA0 ]: Kcur-9 (reshaped) ( 8K) [CUDA0 ]
node #338 ( MUL): Kcur_normed-9 ( 8K) [CUDA0 ]: norm-9 ( 8K) [CUDA0 ] blk.9.attn_k_norm.we ( 1K) [CUDA0 ]
node #339 ( ROPE): Kcur-9 ( 8K) [CUDA0 ]: Kcur_normed-9 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #340 ( MUL_MAT): Vcur-9 ( 8K) [CUDA0 ]: blk.9.attn_v.weight ( 4M) [CUDA0 ] attn_norm-9 ( 15K) [CUDA0 ]
node #342 ( CPY): k_cache_view-9 (copy ( 2K) [CUDA0 ]: Kcur-9 ( 8K) [CUDA0 ] k_cache_view-9 ( 2K) [CUDA0 ]
node #344 ( CPY): v_cache_view-9 (copy ( 2K) [CUDA0 ]: Vcur-9 ( 8K) [CUDA0 ] v_cache_view-9 ( 2K) [CUDA0 ]

SPLIT #20: CPU # 3 inputs: [q-9 ( 16K)] [k-9 ( 544K)] [v-9 ( 544K)]

node #348 (FLASH_ATTN): node_348 ( 16K) [ CPU ]: CPU#q-9#0 ( 16K) [ NULL ] CPU#k-9#0 ( 544K) [ NULL ] CPU#v-9#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #21: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #350 ( MUL_MAT): kqv_out-9 ( 15K) [CUDA0 ]: blk.9.attn_output.we ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #351 ( RMS_NORM): norm-9 ( 15K) [CUDA0 ]: kqv_out-9 ( 15K) [CUDA0 ]
node #352 ( MUL): attn_post_norm-9 ( 15K) [CUDA0 ]: norm-9 ( 15K) [CUDA0 ] blk.9.post_attention ( 15K) [CUDA0 ]
node #353 ( ADD): sa_out-9 ( 15K) [CUDA0 ]: attn_post_norm-9 ( 15K) [CUDA0 ] l_out-8 ( 15K) [CUDA0 ]
node #354 ( RMS_NORM): norm-9 ( 15K) [CUDA0 ]: sa_out-9 ( 15K) [CUDA0 ]
node #355 ( MUL): ffn_norm-9 ( 15K) [CUDA0 ]: norm-9 ( 15K) [CUDA0 ] blk.9.ffn_norm.weigh ( 15K) [CUDA0 ]
node #356 ( MUL_MAT): ffn_gate-9 ( 60K) [CUDA0 ]: blk.9.ffn_gate.weigh ( 31M) [CUDA0 ] ffn_norm-9 ( 15K) [CUDA0 ]
node #357 ( UNARY): ffn_gelu-9 ( 60K) [CUDA0 ]: ffn_gate-9 ( 60K) [CUDA0 ]
node #358 ( MUL_MAT): ffn_up-9 ( 60K) [CUDA0 ]: blk.9.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-9 ( 15K) [CUDA0 ]
node #359 ( MUL): ffn_gate_par-9 ( 60K) [CUDA0 ]: ffn_gelu-9 ( 60K) [CUDA0 ] ffn_up-9 ( 60K) [CUDA0 ]
node #360 ( MUL_MAT): ffn_out-9 ( 15K) [CUDA0 ]: blk.9.ffn_down.weigh ( 31M) [CUDA0 ] ffn_gate_par-9 ( 60K) [CUDA0 ]
node #361 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-9 ( 15K) [CUDA0 ]
node #362 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.9.post_ffw_norm. ( 15K) [CUDA0 ]
node #363 ( ADD): l_out-9 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-9 ( 15K) [CUDA0 ]
node #364 ( RMS_NORM): norm-10 ( 15K) [CUDA0 ]: l_out-9 ( 15K) [CUDA0 ]
node #365 ( MUL): attn_norm-10 ( 15K) [CUDA0 ]: norm-10 ( 15K) [CUDA0 ] blk.10.attn_norm.wei ( 15K) [CUDA0 ]
node #366 ( MUL_MAT): Qcur-10 ( 16K) [CUDA0 ]: blk.10.attn_q.weight ( 8M) [CUDA0 ] attn_norm-10 ( 15K) [CUDA0 ]
node #368 ( RMS_NORM): norm-10 ( 16K) [CUDA0 ]: Qcur-10 (reshaped) ( 16K) [CUDA0 ]
node #369 ( MUL): Qcur_normed-10 ( 16K) [CUDA0 ]: norm-10 ( 16K) [CUDA0 ] blk.10.attn_q_norm.w ( 1K) [CUDA0 ]
node #370 ( ROPE): Qcur-10 ( 16K) [CUDA0 ]: Qcur_normed-10 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #371 ( MUL_MAT): Kcur-10 ( 8K) [CUDA0 ]: blk.10.attn_k.weight ( 4M) [CUDA0 ] attn_norm-10 ( 15K) [CUDA0 ]
node #373 ( RMS_NORM): norm-10 ( 8K) [CUDA0 ]: Kcur-10 (reshaped) ( 8K) [CUDA0 ]
node #374 ( MUL): Kcur_normed-10 ( 8K) [CUDA0 ]: norm-10 ( 8K) [CUDA0 ] blk.10.attn_k_norm.w ( 1K) [CUDA0 ]
node #375 ( ROPE): Kcur-10 ( 8K) [CUDA0 ]: Kcur_normed-10 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #376 ( MUL_MAT): Vcur-10 ( 8K) [CUDA0 ]: blk.10.attn_v.weight ( 4M) [CUDA0 ] attn_norm-10 ( 15K) [CUDA0 ]
node #378 ( CPY): k_cache_view-10 (cop ( 2K) [CUDA0 ]: Kcur-10 ( 8K) [CUDA0 ] k_cache_view-10 ( 2K) [CUDA0 ]
node #380 ( CPY): v_cache_view-10 (cop ( 2K) [CUDA0 ]: Vcur-10 ( 8K) [CUDA0 ] v_cache_view-10 ( 2K) [CUDA0 ]

SPLIT #22: CPU # 3 inputs: [q-10 ( 16K)] [k-10 ( 544K)] [v-10 ( 544K)]

node #384 (FLASH_ATTN): node_384 ( 16K) [ CPU ]: CPU#q-10#0 ( 16K) [ NULL ] CPU#k-10#0 ( 544K) [ NULL ] CPU#v-10#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #23: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #386 ( MUL_MAT): kqv_out-10 ( 15K) [CUDA0 ]: blk.10.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #387 ( RMS_NORM): norm-10 ( 15K) [CUDA0 ]: kqv_out-10 ( 15K) [CUDA0 ]
node #388 ( MUL): attn_post_norm-10 ( 15K) [CUDA0 ]: norm-10 ( 15K) [CUDA0 ] blk.10.post_attentio ( 15K) [CUDA0 ]
node #389 ( ADD): sa_out-10 ( 15K) [CUDA0 ]: attn_post_norm-10 ( 15K) [CUDA0 ] l_out-9 ( 15K) [CUDA0 ]
node #390 ( RMS_NORM): norm-10 ( 15K) [CUDA0 ]: sa_out-10 ( 15K) [CUDA0 ]
node #391 ( MUL): ffn_norm-10 ( 15K) [CUDA0 ]: norm-10 ( 15K) [CUDA0 ] blk.10.ffn_norm.weig ( 15K) [CUDA0 ]
node #392 ( MUL_MAT): ffn_gate-10 ( 60K) [CUDA0 ]: blk.10.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-10 ( 15K) [CUDA0 ]
node #393 ( UNARY): ffn_gelu-10 ( 60K) [CUDA0 ]: ffn_gate-10 ( 60K) [CUDA0 ]
node #394 ( MUL_MAT): ffn_up-10 ( 60K) [CUDA0 ]: blk.10.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-10 ( 15K) [CUDA0 ]
node #395 ( MUL): ffn_gate_par-10 ( 60K) [CUDA0 ]: ffn_gelu-10 ( 60K) [CUDA0 ] ffn_up-10 ( 60K) [CUDA0 ]
node #396 ( MUL_MAT): ffn_out-10 ( 15K) [CUDA0 ]: blk.10.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-10 ( 60K) [CUDA0 ]
node #397 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-10 ( 15K) [CUDA0 ]
node #398 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.10.post_ffw_norm ( 15K) [CUDA0 ]
node #399 ( ADD): l_out-10 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-10 ( 15K) [CUDA0 ]
node #400 ( RMS_NORM): norm-11 ( 15K) [CUDA0 ]: l_out-10 ( 15K) [CUDA0 ]
node #401 ( MUL): attn_norm-11 ( 15K) [CUDA0 ]: norm-11 ( 15K) [CUDA0 ] blk.11.attn_norm.wei ( 15K) [CUDA0 ]
node #402 ( MUL_MAT): Qcur-11 ( 16K) [CUDA0 ]: blk.11.attn_q.weight ( 8M) [CUDA0 ] attn_norm-11 ( 15K) [CUDA0 ]
node #404 ( RMS_NORM): norm-11 ( 16K) [CUDA0 ]: Qcur-11 (reshaped) ( 16K) [CUDA0 ]
node #405 ( MUL): Qcur_normed-11 ( 16K) [CUDA0 ]: norm-11 ( 16K) [CUDA0 ] blk.11.attn_q_norm.w ( 1K) [CUDA0 ]
node #406 ( ROPE): Qcur-11 ( 16K) [CUDA0 ]: Qcur_normed-11 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #407 ( MUL_MAT): Kcur-11 ( 8K) [CUDA0 ]: blk.11.attn_k.weight ( 4M) [CUDA0 ] attn_norm-11 ( 15K) [CUDA0 ]
node #409 ( RMS_NORM): norm-11 ( 8K) [CUDA0 ]: Kcur-11 (reshaped) ( 8K) [CUDA0 ]
node #410 ( MUL): Kcur_normed-11 ( 8K) [CUDA0 ]: norm-11 ( 8K) [CUDA0 ] blk.11.attn_k_norm.w ( 1K) [CUDA0 ]
node #411 ( ROPE): Kcur-11 ( 8K) [CUDA0 ]: Kcur_normed-11 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #412 ( MUL_MAT): Vcur-11 ( 8K) [CUDA0 ]: blk.11.attn_v.weight ( 6M) [CUDA0 ] attn_norm-11 ( 15K) [CUDA0 ]
node #414 ( CPY): k_cache_view-11 (cop ( 2K) [CUDA0 ]: Kcur-11 ( 8K) [CUDA0 ] k_cache_view-11 ( 2K) [CUDA0 ]
node #416 ( CPY): v_cache_view-11 (cop ( 2K) [CUDA0 ]: Vcur-11 ( 8K) [CUDA0 ] v_cache_view-11 ( 2K) [CUDA0 ]

SPLIT #24: CPU # 3 inputs: [q-11 ( 16K)] [k-11 ( 544K)] [v-11 ( 544K)]

node #420 (FLASH_ATTN): node_420 ( 16K) [ CPU ]: CPU#q-11#0 ( 16K) [ NULL ] CPU#k-11#0 ( 544K) [ NULL ] CPU#v-11#0 ( 544K) [ NULL ] CPU#KQ_mask (copy)#0 ( 32K) [ NULL ]

SPLIT #25: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #422 ( MUL_MAT): kqv_out-11 ( 15K) [CUDA0 ]: blk.11.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #423 ( RMS_NORM): norm-11 ( 15K) [CUDA0 ]: kqv_out-11 ( 15K) [CUDA0 ]
node #424 ( MUL): attn_post_norm-11 ( 15K) [CUDA0 ]: norm-11 ( 15K) [CUDA0 ] blk.11.post_attentio ( 15K) [CUDA0 ]
node #425 ( ADD): sa_out-11 ( 15K) [CUDA0 ]: attn_post_norm-11 ( 15K) [CUDA0 ] l_out-10 ( 15K) [CUDA0 ]
node #426 ( RMS_NORM): norm-11 ( 15K) [CUDA0 ]: sa_out-11 ( 15K) [CUDA0 ]
node #427 ( MUL): ffn_norm-11 ( 15K) [CUDA0 ]: norm-11 ( 15K) [CUDA0 ] blk.11.ffn_norm.weig ( 15K) [CUDA0 ]
node #428 ( MUL_MAT): ffn_gate-11 ( 60K) [CUDA0 ]: blk.11.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-11 ( 15K) [CUDA0 ]
node #429 ( UNARY): ffn_gelu-11 ( 60K) [CUDA0 ]: ffn_gate-11 ( 60K) [CUDA0 ]
node #430 ( MUL_MAT): ffn_up-11 ( 60K) [CUDA0 ]: blk.11.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-11 ( 15K) [CUDA0 ]
node #431 ( MUL): ffn_gate_par-11 ( 60K) [CUDA0 ]: ffn_gelu-11 ( 60K) [CUDA0 ] ffn_up-11 ( 60K) [CUDA0 ]
node #432 ( MUL_MAT): ffn_out-11 ( 15K) [CUDA0 ]: blk.11.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-11 ( 60K) [CUDA0 ]
node #433 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-11 ( 15K) [CUDA0 ]
node #434 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.11.post_ffw_norm ( 15K) [CUDA0 ]
node #435 ( ADD): l_out-11 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-11 ( 15K) [CUDA0 ]
node #436 ( RMS_NORM): norm-12 ( 15K) [CUDA0 ]: l_out-11 ( 15K) [CUDA0 ]
node #437 ( MUL): attn_norm-12 ( 15K) [CUDA0 ]: norm-12 ( 15K) [CUDA0 ] blk.12.attn_norm.wei ( 15K) [CUDA0 ]
node #438 ( MUL_MAT): Qcur-12 ( 16K) [CUDA0 ]: blk.12.attn_q.weight ( 8M) [CUDA0 ] attn_norm-12 ( 15K) [CUDA0 ]
node #440 ( RMS_NORM): norm-12 ( 16K) [CUDA0 ]: Qcur-12 (reshaped) ( 16K) [CUDA0 ]
node #441 ( MUL): Qcur_normed-12 ( 16K) [CUDA0 ]: norm-12 ( 16K) [CUDA0 ] blk.12.attn_q_norm.w ( 1K) [CUDA0 ]
node #442 ( ROPE): Qcur-12 ( 16K) [CUDA0 ]: Qcur_normed-12 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #443 ( MUL_MAT): Kcur-12 ( 8K) [CUDA0 ]: blk.12.attn_k.weight ( 4M) [CUDA0 ] attn_norm-12 ( 15K) [CUDA0 ]
node #445 ( RMS_NORM): norm-12 ( 8K) [CUDA0 ]: Kcur-12 (reshaped) ( 8K) [CUDA0 ]
node #446 ( MUL): Kcur_normed-12 ( 8K) [CUDA0 ]: norm-12 ( 8K) [CUDA0 ] blk.12.attn_k_norm.w ( 1K) [CUDA0 ]
node #447 ( ROPE): Kcur-12 ( 8K) [CUDA0 ]: Kcur_normed-12 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #448 ( MUL_MAT): Vcur-12 ( 8K) [CUDA0 ]: blk.12.attn_v.weight ( 4M) [CUDA0 ] attn_norm-12 ( 15K) [CUDA0 ]
node #450 ( CPY): k_cache_view-12 (cop ( 2K) [CUDA0 ]: Kcur-12 ( 8K) [CUDA0 ] k_cache_view-12 ( 2K) [CUDA0 ]
node #452 ( CPY): v_cache_view-12 (cop ( 2K) [CUDA0 ]: Vcur-12 ( 8K) [CUDA0 ] v_cache_view-12 ( 2K) [CUDA0 ]

SPLIT #26: CPU # 3 inputs: [q-12 ( 16K)] [k-12 ( 544K)] [v-12 ( 544K)]

node #456 (FLASH_ATTN): node_456 ( 16K) [ CPU ]: CPU#q-12#0 ( 16K) [ NULL ] CPU#k-12#0 ( 544K) [ NULL ] CPU#v-12#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #27: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #458 ( MUL_MAT): kqv_out-12 ( 15K) [CUDA0 ]: blk.12.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #459 ( RMS_NORM): norm-12 ( 15K) [CUDA0 ]: kqv_out-12 ( 15K) [CUDA0 ]
node #460 ( MUL): attn_post_norm-12 ( 15K) [CUDA0 ]: norm-12 ( 15K) [CUDA0 ] blk.12.post_attentio ( 15K) [CUDA0 ]
node #461 ( ADD): sa_out-12 ( 15K) [CUDA0 ]: attn_post_norm-12 ( 15K) [CUDA0 ] l_out-11 ( 15K) [CUDA0 ]
node #462 ( RMS_NORM): norm-12 ( 15K) [CUDA0 ]: sa_out-12 ( 15K) [CUDA0 ]
node #463 ( MUL): ffn_norm-12 ( 15K) [CUDA0 ]: norm-12 ( 15K) [CUDA0 ] blk.12.ffn_norm.weig ( 15K) [CUDA0 ]
node #464 ( MUL_MAT): ffn_gate-12 ( 60K) [CUDA0 ]: blk.12.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-12 ( 15K) [CUDA0 ]
node #465 ( UNARY): ffn_gelu-12 ( 60K) [CUDA0 ]: ffn_gate-12 ( 60K) [CUDA0 ]
node #466 ( MUL_MAT): ffn_up-12 ( 60K) [CUDA0 ]: blk.12.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-12 ( 15K) [CUDA0 ]
node #467 ( MUL): ffn_gate_par-12 ( 60K) [CUDA0 ]: ffn_gelu-12 ( 60K) [CUDA0 ] ffn_up-12 ( 60K) [CUDA0 ]
node #468 ( MUL_MAT): ffn_out-12 ( 15K) [CUDA0 ]: blk.12.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-12 ( 60K) [CUDA0 ]
node #469 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-12 ( 15K) [CUDA0 ]
node #470 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.12.post_ffw_norm ( 15K) [CUDA0 ]
node #471 ( ADD): l_out-12 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-12 ( 15K) [CUDA0 ]
node #472 ( RMS_NORM): norm-13 ( 15K) [CUDA0 ]: l_out-12 ( 15K) [CUDA0 ]
node #473 ( MUL): attn_norm-13 ( 15K) [CUDA0 ]: norm-13 ( 15K) [CUDA0 ] blk.13.attn_norm.wei ( 15K) [CUDA0 ]
node #474 ( MUL_MAT): Qcur-13 ( 16K) [CUDA0 ]: blk.13.attn_q.weight ( 8M) [CUDA0 ] attn_norm-13 ( 15K) [CUDA0 ]
node #476 ( RMS_NORM): norm-13 ( 16K) [CUDA0 ]: Qcur-13 (reshaped) ( 16K) [CUDA0 ]
node #477 ( MUL): Qcur_normed-13 ( 16K) [CUDA0 ]: norm-13 ( 16K) [CUDA0 ] blk.13.attn_q_norm.w ( 1K) [CUDA0 ]
node #478 ( ROPE): Qcur-13 ( 16K) [CUDA0 ]: Qcur_normed-13 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #479 ( MUL_MAT): Kcur-13 ( 8K) [CUDA0 ]: blk.13.attn_k.weight ( 4M) [CUDA0 ] attn_norm-13 ( 15K) [CUDA0 ]
node #481 ( RMS_NORM): norm-13 ( 8K) [CUDA0 ]: Kcur-13 (reshaped) ( 8K) [CUDA0 ]
node #482 ( MUL): Kcur_normed-13 ( 8K) [CUDA0 ]: norm-13 ( 8K) [CUDA0 ] blk.13.attn_k_norm.w ( 1K) [CUDA0 ]
node #483 ( ROPE): Kcur-13 ( 8K) [CUDA0 ]: Kcur_normed-13 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #484 ( MUL_MAT): Vcur-13 ( 8K) [CUDA0 ]: blk.13.attn_v.weight ( 4M) [CUDA0 ] attn_norm-13 ( 15K) [CUDA0 ]
node #486 ( CPY): k_cache_view-13 (cop ( 2K) [CUDA0 ]: Kcur-13 ( 8K) [CUDA0 ] k_cache_view-13 ( 2K) [CUDA0 ]
node #488 ( CPY): v_cache_view-13 (cop ( 2K) [CUDA0 ]: Vcur-13 ( 8K) [CUDA0 ] v_cache_view-13 ( 2K) [CUDA0 ]

SPLIT #28: CPU # 3 inputs: [q-13 ( 16K)] [k-13 ( 544K)] [v-13 ( 544K)]

node #492 (FLASH_ATTN): node_492 ( 16K) [ CPU ]: CPU#q-13#0 ( 16K) [ NULL ] CPU#k-13#0 ( 544K) [ NULL ] CPU#v-13#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #29: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #494 ( MUL_MAT): kqv_out-13 ( 15K) [CUDA0 ]: blk.13.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #495 ( RMS_NORM): norm-13 ( 15K) [CUDA0 ]: kqv_out-13 ( 15K) [CUDA0 ]
node #496 ( MUL): attn_post_norm-13 ( 15K) [CUDA0 ]: norm-13 ( 15K) [CUDA0 ] blk.13.post_attentio ( 15K) [CUDA0 ]
node #497 ( ADD): sa_out-13 ( 15K) [CUDA0 ]: attn_post_norm-13 ( 15K) [CUDA0 ] l_out-12 ( 15K) [CUDA0 ]
node #498 ( RMS_NORM): norm-13 ( 15K) [CUDA0 ]: sa_out-13 ( 15K) [CUDA0 ]
node #499 ( MUL): ffn_norm-13 ( 15K) [CUDA0 ]: norm-13 ( 15K) [CUDA0 ] blk.13.ffn_norm.weig ( 15K) [CUDA0 ]
node #500 ( MUL_MAT): ffn_gate-13 ( 60K) [CUDA0 ]: blk.13.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-13 ( 15K) [CUDA0 ]
node #501 ( UNARY): ffn_gelu-13 ( 60K) [CUDA0 ]: ffn_gate-13 ( 60K) [CUDA0 ]
node #502 ( MUL_MAT): ffn_up-13 ( 60K) [CUDA0 ]: blk.13.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-13 ( 15K) [CUDA0 ]
node #503 ( MUL): ffn_gate_par-13 ( 60K) [CUDA0 ]: ffn_gelu-13 ( 60K) [CUDA0 ] ffn_up-13 ( 60K) [CUDA0 ]
node #504 ( MUL_MAT): ffn_out-13 ( 15K) [CUDA0 ]: blk.13.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-13 ( 60K) [CUDA0 ]
node #505 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-13 ( 15K) [CUDA0 ]
node #506 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.13.post_ffw_norm ( 15K) [CUDA0 ]
node #507 ( ADD): l_out-13 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-13 ( 15K) [CUDA0 ]
node #508 ( RMS_NORM): norm-14 ( 15K) [CUDA0 ]: l_out-13 ( 15K) [CUDA0 ]
node #509 ( MUL): attn_norm-14 ( 15K) [CUDA0 ]: norm-14 ( 15K) [CUDA0 ] blk.14.attn_norm.wei ( 15K) [CUDA0 ]
node #510 ( MUL_MAT): Qcur-14 ( 16K) [CUDA0 ]: blk.14.attn_q.weight ( 8M) [CUDA0 ] attn_norm-14 ( 15K) [CUDA0 ]
node #512 ( RMS_NORM): norm-14 ( 16K) [CUDA0 ]: Qcur-14 (reshaped) ( 16K) [CUDA0 ]
node #513 ( MUL): Qcur_normed-14 ( 16K) [CUDA0 ]: norm-14 ( 16K) [CUDA0 ] blk.14.attn_q_norm.w ( 1K) [CUDA0 ]
node #514 ( ROPE): Qcur-14 ( 16K) [CUDA0 ]: Qcur_normed-14 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #515 ( MUL_MAT): Kcur-14 ( 8K) [CUDA0 ]: blk.14.attn_k.weight ( 4M) [CUDA0 ] attn_norm-14 ( 15K) [CUDA0 ]
node #517 ( RMS_NORM): norm-14 ( 8K) [CUDA0 ]: Kcur-14 (reshaped) ( 8K) [CUDA0 ]
node #518 ( MUL): Kcur_normed-14 ( 8K) [CUDA0 ]: norm-14 ( 8K) [CUDA0 ] blk.14.attn_k_norm.w ( 1K) [CUDA0 ]
node #519 ( ROPE): Kcur-14 ( 8K) [CUDA0 ]: Kcur_normed-14 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #520 ( MUL_MAT): Vcur-14 ( 8K) [CUDA0 ]: blk.14.attn_v.weight ( 6M) [CUDA0 ] attn_norm-14 ( 15K) [CUDA0 ]
node #522 ( CPY): k_cache_view-14 (cop ( 2K) [CUDA0 ]: Kcur-14 ( 8K) [CUDA0 ] k_cache_view-14 ( 2K) [CUDA0 ]
node #524 ( CPY): v_cache_view-14 (cop ( 2K) [CUDA0 ]: Vcur-14 ( 8K) [CUDA0 ] v_cache_view-14 ( 2K) [CUDA0 ]

SPLIT #30: CPU # 3 inputs: [q-14 ( 16K)] [k-14 ( 544K)] [v-14 ( 544K)]

node #528 (FLASH_ATTN): node_528 ( 16K) [ CPU ]: CPU#q-14#0 ( 16K) [ NULL ] CPU#k-14#0 ( 544K) [ NULL ] CPU#v-14#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #31: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #530 ( MUL_MAT): kqv_out-14 ( 15K) [CUDA0 ]: blk.14.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #531 ( RMS_NORM): norm-14 ( 15K) [CUDA0 ]: kqv_out-14 ( 15K) [CUDA0 ]
node #532 ( MUL): attn_post_norm-14 ( 15K) [CUDA0 ]: norm-14 ( 15K) [CUDA0 ] blk.14.post_attentio ( 15K) [CUDA0 ]
node #533 ( ADD): sa_out-14 ( 15K) [CUDA0 ]: attn_post_norm-14 ( 15K) [CUDA0 ] l_out-13 ( 15K) [CUDA0 ]
node #534 ( RMS_NORM): norm-14 ( 15K) [CUDA0 ]: sa_out-14 ( 15K) [CUDA0 ]
node #535 ( MUL): ffn_norm-14 ( 15K) [CUDA0 ]: norm-14 ( 15K) [CUDA0 ] blk.14.ffn_norm.weig ( 15K) [CUDA0 ]
node #536 ( MUL_MAT): ffn_gate-14 ( 60K) [CUDA0 ]: blk.14.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-14 ( 15K) [CUDA0 ]
node #537 ( UNARY): ffn_gelu-14 ( 60K) [CUDA0 ]: ffn_gate-14 ( 60K) [CUDA0 ]
node #538 ( MUL_MAT): ffn_up-14 ( 60K) [CUDA0 ]: blk.14.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-14 ( 15K) [CUDA0 ]
node #539 ( MUL): ffn_gate_par-14 ( 60K) [CUDA0 ]: ffn_gelu-14 ( 60K) [CUDA0 ] ffn_up-14 ( 60K) [CUDA0 ]
node #540 ( MUL_MAT): ffn_out-14 ( 15K) [CUDA0 ]: blk.14.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-14 ( 60K) [CUDA0 ]
node #541 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-14 ( 15K) [CUDA0 ]
node #542 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.14.post_ffw_norm ( 15K) [CUDA0 ]
node #543 ( ADD): l_out-14 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-14 ( 15K) [CUDA0 ]
node #544 ( RMS_NORM): norm-15 ( 15K) [CUDA0 ]: l_out-14 ( 15K) [CUDA0 ]
node #545 ( MUL): attn_norm-15 ( 15K) [CUDA0 ]: norm-15 ( 15K) [CUDA0 ] blk.15.attn_norm.wei ( 15K) [CUDA0 ]
node #546 ( MUL_MAT): Qcur-15 ( 16K) [CUDA0 ]: blk.15.attn_q.weight ( 8M) [CUDA0 ] attn_norm-15 ( 15K) [CUDA0 ]
node #548 ( RMS_NORM): norm-15 ( 16K) [CUDA0 ]: Qcur-15 (reshaped) ( 16K) [CUDA0 ]
node #549 ( MUL): Qcur_normed-15 ( 16K) [CUDA0 ]: norm-15 ( 16K) [CUDA0 ] blk.15.attn_q_norm.w ( 1K) [CUDA0 ]
node #550 ( ROPE): Qcur-15 ( 16K) [CUDA0 ]: Qcur_normed-15 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #551 ( MUL_MAT): Kcur-15 ( 8K) [CUDA0 ]: blk.15.attn_k.weight ( 4M) [CUDA0 ] attn_norm-15 ( 15K) [CUDA0 ]
node #553 ( RMS_NORM): norm-15 ( 8K) [CUDA0 ]: Kcur-15 (reshaped) ( 8K) [CUDA0 ]
node #554 ( MUL): Kcur_normed-15 ( 8K) [CUDA0 ]: norm-15 ( 8K) [CUDA0 ] blk.15.attn_k_norm.w ( 1K) [CUDA0 ]
node #555 ( ROPE): Kcur-15 ( 8K) [CUDA0 ]: Kcur_normed-15 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #556 ( MUL_MAT): Vcur-15 ( 8K) [CUDA0 ]: blk.15.attn_v.weight ( 4M) [CUDA0 ] attn_norm-15 ( 15K) [CUDA0 ]
node #558 ( CPY): k_cache_view-15 (cop ( 2K) [CUDA0 ]: Kcur-15 ( 8K) [CUDA0 ] k_cache_view-15 ( 2K) [CUDA0 ]
node #560 ( CPY): v_cache_view-15 (cop ( 2K) [CUDA0 ]: Vcur-15 ( 8K) [CUDA0 ] v_cache_view-15 ( 2K) [CUDA0 ]

SPLIT #32: CPU # 3 inputs: [q-15 ( 16K)] [k-15 ( 544K)] [v-15 ( 544K)]

node #564 (FLASH_ATTN): node_564 ( 16K) [ CPU ]: CPU#q-15#0 ( 16K) [ NULL ] CPU#k-15#0 ( 544K) [ NULL ] CPU#v-15#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #33: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #566 ( MUL_MAT): kqv_out-15 ( 15K) [CUDA0 ]: blk.15.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #567 ( RMS_NORM): norm-15 ( 15K) [CUDA0 ]: kqv_out-15 ( 15K) [CUDA0 ]
node #568 ( MUL): attn_post_norm-15 ( 15K) [CUDA0 ]: norm-15 ( 15K) [CUDA0 ] blk.15.post_attentio ( 15K) [CUDA0 ]
node #569 ( ADD): sa_out-15 ( 15K) [CUDA0 ]: attn_post_norm-15 ( 15K) [CUDA0 ] l_out-14 ( 15K) [CUDA0 ]
node #570 ( RMS_NORM): norm-15 ( 15K) [CUDA0 ]: sa_out-15 ( 15K) [CUDA0 ]
node #571 ( MUL): ffn_norm-15 ( 15K) [CUDA0 ]: norm-15 ( 15K) [CUDA0 ] blk.15.ffn_norm.weig ( 15K) [CUDA0 ]
node #572 ( MUL_MAT): ffn_gate-15 ( 60K) [CUDA0 ]: blk.15.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-15 ( 15K) [CUDA0 ]
node #573 ( UNARY): ffn_gelu-15 ( 60K) [CUDA0 ]: ffn_gate-15 ( 60K) [CUDA0 ]
node #574 ( MUL_MAT): ffn_up-15 ( 60K) [CUDA0 ]: blk.15.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-15 ( 15K) [CUDA0 ]
node #575 ( MUL): ffn_gate_par-15 ( 60K) [CUDA0 ]: ffn_gelu-15 ( 60K) [CUDA0 ] ffn_up-15 ( 60K) [CUDA0 ]
node #576 ( MUL_MAT): ffn_out-15 ( 15K) [CUDA0 ]: blk.15.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-15 ( 60K) [CUDA0 ]
node #577 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-15 ( 15K) [CUDA0 ]
node #578 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.15.post_ffw_norm ( 15K) [CUDA0 ]
node #579 ( ADD): l_out-15 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-15 ( 15K) [CUDA0 ]
node #580 ( RMS_NORM): norm-16 ( 15K) [CUDA0 ]: l_out-15 ( 15K) [CUDA0 ]
node #581 ( MUL): attn_norm-16 ( 15K) [CUDA0 ]: norm-16 ( 15K) [CUDA0 ] blk.16.attn_norm.wei ( 15K) [CUDA0 ]
node #582 ( MUL_MAT): Qcur-16 ( 16K) [CUDA0 ]: blk.16.attn_q.weight ( 8M) [CUDA0 ] attn_norm-16 ( 15K) [CUDA0 ]
node #584 ( RMS_NORM): norm-16 ( 16K) [CUDA0 ]: Qcur-16 (reshaped) ( 16K) [CUDA0 ]
node #585 ( MUL): Qcur_normed-16 ( 16K) [CUDA0 ]: norm-16 ( 16K) [CUDA0 ] blk.16.attn_q_norm.w ( 1K) [CUDA0 ]
node #586 ( ROPE): Qcur-16 ( 16K) [CUDA0 ]: Qcur_normed-16 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #587 ( MUL_MAT): Kcur-16 ( 8K) [CUDA0 ]: blk.16.attn_k.weight ( 4M) [CUDA0 ] attn_norm-16 ( 15K) [CUDA0 ]
node #589 ( RMS_NORM): norm-16 ( 8K) [CUDA0 ]: Kcur-16 (reshaped) ( 8K) [CUDA0 ]
node #590 ( MUL): Kcur_normed-16 ( 8K) [CUDA0 ]: norm-16 ( 8K) [CUDA0 ] blk.16.attn_k_norm.w ( 1K) [CUDA0 ]
node #591 ( ROPE): Kcur-16 ( 8K) [CUDA0 ]: Kcur_normed-16 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #592 ( MUL_MAT): Vcur-16 ( 8K) [CUDA0 ]: blk.16.attn_v.weight ( 4M) [CUDA0 ] attn_norm-16 ( 15K) [CUDA0 ]
node #594 ( CPY): k_cache_view-16 (cop ( 2K) [CUDA0 ]: Kcur-16 ( 8K) [CUDA0 ] k_cache_view-16 ( 2K) [CUDA0 ]
node #596 ( CPY): v_cache_view-16 (cop ( 2K) [CUDA0 ]: Vcur-16 ( 8K) [CUDA0 ] v_cache_view-16 ( 2K) [CUDA0 ]

SPLIT #34: CPU # 3 inputs: [q-16 ( 16K)] [k-16 ( 544K)] [v-16 ( 544K)]

node #600 (FLASH_ATTN): node_600 ( 16K) [ CPU ]: CPU#q-16#0 ( 16K) [ NULL ] CPU#k-16#0 ( 544K) [ NULL ] CPU#v-16#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #35: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #602 ( MUL_MAT): kqv_out-16 ( 15K) [CUDA0 ]: blk.16.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #603 ( RMS_NORM): norm-16 ( 15K) [CUDA0 ]: kqv_out-16 ( 15K) [CUDA0 ]
node #604 ( MUL): attn_post_norm-16 ( 15K) [CUDA0 ]: norm-16 ( 15K) [CUDA0 ] blk.16.post_attentio ( 15K) [CUDA0 ]
node #605 ( ADD): sa_out-16 ( 15K) [CUDA0 ]: attn_post_norm-16 ( 15K) [CUDA0 ] l_out-15 ( 15K) [CUDA0 ]
node #606 ( RMS_NORM): norm-16 ( 15K) [CUDA0 ]: sa_out-16 ( 15K) [CUDA0 ]
node #607 ( MUL): ffn_norm-16 ( 15K) [CUDA0 ]: norm-16 ( 15K) [CUDA0 ] blk.16.ffn_norm.weig ( 15K) [CUDA0 ]
node #608 ( MUL_MAT): ffn_gate-16 ( 60K) [CUDA0 ]: blk.16.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-16 ( 15K) [CUDA0 ]
node #609 ( UNARY): ffn_gelu-16 ( 60K) [CUDA0 ]: ffn_gate-16 ( 60K) [CUDA0 ]
node #610 ( MUL_MAT): ffn_up-16 ( 60K) [CUDA0 ]: blk.16.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-16 ( 15K) [CUDA0 ]
node #611 ( MUL): ffn_gate_par-16 ( 60K) [CUDA0 ]: ffn_gelu-16 ( 60K) [CUDA0 ] ffn_up-16 ( 60K) [CUDA0 ]
node #612 ( MUL_MAT): ffn_out-16 ( 15K) [CUDA0 ]: blk.16.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-16 ( 60K) [CUDA0 ]
node #613 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-16 ( 15K) [CUDA0 ]
node #614 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.16.post_ffw_norm ( 15K) [CUDA0 ]
node #615 ( ADD): l_out-16 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-16 ( 15K) [CUDA0 ]
node #616 ( RMS_NORM): norm-17 ( 15K) [CUDA0 ]: l_out-16 ( 15K) [CUDA0 ]
node #617 ( MUL): attn_norm-17 ( 15K) [CUDA0 ]: norm-17 ( 15K) [CUDA0 ] blk.17.attn_norm.wei ( 15K) [CUDA0 ]
node #618 ( MUL_MAT): Qcur-17 ( 16K) [CUDA0 ]: blk.17.attn_q.weight ( 8M) [CUDA0 ] attn_norm-17 ( 15K) [CUDA0 ]
node #620 ( RMS_NORM): norm-17 ( 16K) [CUDA0 ]: Qcur-17 (reshaped) ( 16K) [CUDA0 ]
node #621 ( MUL): Qcur_normed-17 ( 16K) [CUDA0 ]: norm-17 ( 16K) [CUDA0 ] blk.17.attn_q_norm.w ( 1K) [CUDA0 ]
node #622 ( ROPE): Qcur-17 ( 16K) [CUDA0 ]: Qcur_normed-17 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #623 ( MUL_MAT): Kcur-17 ( 8K) [CUDA0 ]: blk.17.attn_k.weight ( 4M) [CUDA0 ] attn_norm-17 ( 15K) [CUDA0 ]
node #625 ( RMS_NORM): norm-17 ( 8K) [CUDA0 ]: Kcur-17 (reshaped) ( 8K) [CUDA0 ]
node #626 ( MUL): Kcur_normed-17 ( 8K) [CUDA0 ]: norm-17 ( 8K) [CUDA0 ] blk.17.attn_k_norm.w ( 1K) [CUDA0 ]
node #627 ( ROPE): Kcur-17 ( 8K) [CUDA0 ]: Kcur_normed-17 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #628 ( MUL_MAT): Vcur-17 ( 8K) [CUDA0 ]: blk.17.attn_v.weight ( 6M) [CUDA0 ] attn_norm-17 ( 15K) [CUDA0 ]
node #630 ( CPY): k_cache_view-17 (cop ( 2K) [CUDA0 ]: Kcur-17 ( 8K) [CUDA0 ] k_cache_view-17 ( 2K) [CUDA0 ]
node #632 ( CPY): v_cache_view-17 (cop ( 2K) [CUDA0 ]: Vcur-17 ( 8K) [CUDA0 ] v_cache_view-17 ( 2K) [CUDA0 ]

SPLIT #36: CPU # 3 inputs: [q-17 ( 16K)] [k-17 ( 544K)] [v-17 ( 544K)]

node #636 (FLASH_ATTN): node_636 ( 16K) [ CPU ]: CPU#q-17#0 ( 16K) [ NULL ] CPU#k-17#0 ( 544K) [ NULL ] CPU#v-17#0 ( 544K) [ NULL ] CPU#KQ_mask (copy)#0 ( 32K) [ NULL ]

SPLIT #37: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #638 ( MUL_MAT): kqv_out-17 ( 15K) [CUDA0 ]: blk.17.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #639 ( RMS_NORM): norm-17 ( 15K) [CUDA0 ]: kqv_out-17 ( 15K) [CUDA0 ]
node #640 ( MUL): attn_post_norm-17 ( 15K) [CUDA0 ]: norm-17 ( 15K) [CUDA0 ] blk.17.post_attentio ( 15K) [CUDA0 ]
node #641 ( ADD): sa_out-17 ( 15K) [CUDA0 ]: attn_post_norm-17 ( 15K) [CUDA0 ] l_out-16 ( 15K) [CUDA0 ]
node #642 ( RMS_NORM): norm-17 ( 15K) [CUDA0 ]: sa_out-17 ( 15K) [CUDA0 ]
node #643 ( MUL): ffn_norm-17 ( 15K) [CUDA0 ]: norm-17 ( 15K) [CUDA0 ] blk.17.ffn_norm.weig ( 15K) [CUDA0 ]
node #644 ( MUL_MAT): ffn_gate-17 ( 60K) [CUDA0 ]: blk.17.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-17 ( 15K) [CUDA0 ]
node #645 ( UNARY): ffn_gelu-17 ( 60K) [CUDA0 ]: ffn_gate-17 ( 60K) [CUDA0 ]
node #646 ( MUL_MAT): ffn_up-17 ( 60K) [CUDA0 ]: blk.17.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-17 ( 15K) [CUDA0 ]
node #647 ( MUL): ffn_gate_par-17 ( 60K) [CUDA0 ]: ffn_gelu-17 ( 60K) [CUDA0 ] ffn_up-17 ( 60K) [CUDA0 ]
node #648 ( MUL_MAT): ffn_out-17 ( 15K) [CUDA0 ]: blk.17.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-17 ( 60K) [CUDA0 ]
node #649 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-17 ( 15K) [CUDA0 ]
node #650 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.17.post_ffw_norm ( 15K) [CUDA0 ]
node #651 ( ADD): l_out-17 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-17 ( 15K) [CUDA0 ]
node #652 ( RMS_NORM): norm-18 ( 15K) [CUDA0 ]: l_out-17 ( 15K) [CUDA0 ]
node #653 ( MUL): attn_norm-18 ( 15K) [CUDA0 ]: norm-18 ( 15K) [CUDA0 ] blk.18.attn_norm.wei ( 15K) [CUDA0 ]
node #654 ( MUL_MAT): Qcur-18 ( 16K) [CUDA0 ]: blk.18.attn_q.weight ( 8M) [CUDA0 ] attn_norm-18 ( 15K) [CUDA0 ]
node #656 ( RMS_NORM): norm-18 ( 16K) [CUDA0 ]: Qcur-18 (reshaped) ( 16K) [CUDA0 ]
node #657 ( MUL): Qcur_normed-18 ( 16K) [CUDA0 ]: norm-18 ( 16K) [CUDA0 ] blk.18.attn_q_norm.w ( 1K) [CUDA0 ]
node #658 ( ROPE): Qcur-18 ( 16K) [CUDA0 ]: Qcur_normed-18 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #659 ( MUL_MAT): Kcur-18 ( 8K) [CUDA0 ]: blk.18.attn_k.weight ( 4M) [CUDA0 ] attn_norm-18 ( 15K) [CUDA0 ]
node #661 ( RMS_NORM): norm-18 ( 8K) [CUDA0 ]: Kcur-18 (reshaped) ( 8K) [CUDA0 ]
node #662 ( MUL): Kcur_normed-18 ( 8K) [CUDA0 ]: norm-18 ( 8K) [CUDA0 ] blk.18.attn_k_norm.w ( 1K) [CUDA0 ]
node #663 ( ROPE): Kcur-18 ( 8K) [CUDA0 ]: Kcur_normed-18 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #664 ( MUL_MAT): Vcur-18 ( 8K) [CUDA0 ]: blk.18.attn_v.weight ( 4M) [CUDA0 ] attn_norm-18 ( 15K) [CUDA0 ]
node #666 ( CPY): k_cache_view-18 (cop ( 2K) [CUDA0 ]: Kcur-18 ( 8K) [CUDA0 ] k_cache_view-18 ( 2K) [CUDA0 ]
node #668 ( CPY): v_cache_view-18 (cop ( 2K) [CUDA0 ]: Vcur-18 ( 8K) [CUDA0 ] v_cache_view-18 ( 2K) [CUDA0 ]

SPLIT #38: CPU # 3 inputs: [q-18 ( 16K)] [k-18 ( 544K)] [v-18 ( 544K)]

node #672 (FLASH_ATTN): node_672 ( 16K) [ CPU ]: CPU#q-18#0 ( 16K) [ NULL ] CPU#k-18#0 ( 544K) [ NULL ] CPU#v-18#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #39: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #674 ( MUL_MAT): kqv_out-18 ( 15K) [CUDA0 ]: blk.18.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #675 ( RMS_NORM): norm-18 ( 15K) [CUDA0 ]: kqv_out-18 ( 15K) [CUDA0 ]
node #676 ( MUL): attn_post_norm-18 ( 15K) [CUDA0 ]: norm-18 ( 15K) [CUDA0 ] blk.18.post_attentio ( 15K) [CUDA0 ]
node #677 ( ADD): sa_out-18 ( 15K) [CUDA0 ]: attn_post_norm-18 ( 15K) [CUDA0 ] l_out-17 ( 15K) [CUDA0 ]
node #678 ( RMS_NORM): norm-18 ( 15K) [CUDA0 ]: sa_out-18 ( 15K) [CUDA0 ]
node #679 ( MUL): ffn_norm-18 ( 15K) [CUDA0 ]: norm-18 ( 15K) [CUDA0 ] blk.18.ffn_norm.weig ( 15K) [CUDA0 ]
node #680 ( MUL_MAT): ffn_gate-18 ( 60K) [CUDA0 ]: blk.18.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-18 ( 15K) [CUDA0 ]
node #681 ( UNARY): ffn_gelu-18 ( 60K) [CUDA0 ]: ffn_gate-18 ( 60K) [CUDA0 ]
node #682 ( MUL_MAT): ffn_up-18 ( 60K) [CUDA0 ]: blk.18.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-18 ( 15K) [CUDA0 ]
node #683 ( MUL): ffn_gate_par-18 ( 60K) [CUDA0 ]: ffn_gelu-18 ( 60K) [CUDA0 ] ffn_up-18 ( 60K) [CUDA0 ]
node #684 ( MUL_MAT): ffn_out-18 ( 15K) [CUDA0 ]: blk.18.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-18 ( 60K) [CUDA0 ]
node #685 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-18 ( 15K) [CUDA0 ]
node #686 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.18.post_ffw_norm ( 15K) [CUDA0 ]
node #687 ( ADD): l_out-18 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-18 ( 15K) [CUDA0 ]
node #688 ( RMS_NORM): norm-19 ( 15K) [CUDA0 ]: l_out-18 ( 15K) [CUDA0 ]
node #689 ( MUL): attn_norm-19 ( 15K) [CUDA0 ]: norm-19 ( 15K) [CUDA0 ] blk.19.attn_norm.wei ( 15K) [CUDA0 ]
node #690 ( MUL_MAT): Qcur-19 ( 16K) [CUDA0 ]: blk.19.attn_q.weight ( 8M) [CUDA0 ] attn_norm-19 ( 15K) [CUDA0 ]
node #692 ( RMS_NORM): norm-19 ( 16K) [CUDA0 ]: Qcur-19 (reshaped) ( 16K) [CUDA0 ]
node #693 ( MUL): Qcur_normed-19 ( 16K) [CUDA0 ]: norm-19 ( 16K) [CUDA0 ] blk.19.attn_q_norm.w ( 1K) [CUDA0 ]
node #694 ( ROPE): Qcur-19 ( 16K) [CUDA0 ]: Qcur_normed-19 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #695 ( MUL_MAT): Kcur-19 ( 8K) [CUDA0 ]: blk.19.attn_k.weight ( 4M) [CUDA0 ] attn_norm-19 ( 15K) [CUDA0 ]
node #697 ( RMS_NORM): norm-19 ( 8K) [CUDA0 ]: Kcur-19 (reshaped) ( 8K) [CUDA0 ]
node #698 ( MUL): Kcur_normed-19 ( 8K) [CUDA0 ]: norm-19 ( 8K) [CUDA0 ] blk.19.attn_k_norm.w ( 1K) [CUDA0 ]
node #699 ( ROPE): Kcur-19 ( 8K) [CUDA0 ]: Kcur_normed-19 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #700 ( MUL_MAT): Vcur-19 ( 8K) [CUDA0 ]: blk.19.attn_v.weight ( 4M) [CUDA0 ] attn_norm-19 ( 15K) [CUDA0 ]
node #702 ( CPY): k_cache_view-19 (cop ( 2K) [CUDA0 ]: Kcur-19 ( 8K) [CUDA0 ] k_cache_view-19 ( 2K) [CUDA0 ]
node #704 ( CPY): v_cache_view-19 (cop ( 2K) [CUDA0 ]: Vcur-19 ( 8K) [CUDA0 ] v_cache_view-19 ( 2K) [CUDA0 ]

SPLIT #40: CPU # 3 inputs: [q-19 ( 16K)] [k-19 ( 544K)] [v-19 ( 544K)]

node #708 (FLASH_ATTN): node_708 ( 16K) [ CPU ]: CPU#q-19#0 ( 16K) [ NULL ] CPU#k-19#0 ( 544K) [ NULL ] CPU#v-19#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #41: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #710 ( MUL_MAT): kqv_out-19 ( 15K) [CUDA0 ]: blk.19.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #711 ( RMS_NORM): norm-19 ( 15K) [CUDA0 ]: kqv_out-19 ( 15K) [CUDA0 ]
node #712 ( MUL): attn_post_norm-19 ( 15K) [CUDA0 ]: norm-19 ( 15K) [CUDA0 ] blk.19.post_attentio ( 15K) [CUDA0 ]
node #713 ( ADD): sa_out-19 ( 15K) [CUDA0 ]: attn_post_norm-19 ( 15K) [CUDA0 ] l_out-18 ( 15K) [CUDA0 ]
node #714 ( RMS_NORM): norm-19 ( 15K) [CUDA0 ]: sa_out-19 ( 15K) [CUDA0 ]
node #715 ( MUL): ffn_norm-19 ( 15K) [CUDA0 ]: norm-19 ( 15K) [CUDA0 ] blk.19.ffn_norm.weig ( 15K) [CUDA0 ]
node #716 ( MUL_MAT): ffn_gate-19 ( 60K) [CUDA0 ]: blk.19.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-19 ( 15K) [CUDA0 ]
node #717 ( UNARY): ffn_gelu-19 ( 60K) [CUDA0 ]: ffn_gate-19 ( 60K) [CUDA0 ]
node #718 ( MUL_MAT): ffn_up-19 ( 60K) [CUDA0 ]: blk.19.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-19 ( 15K) [CUDA0 ]
node #719 ( MUL): ffn_gate_par-19 ( 60K) [CUDA0 ]: ffn_gelu-19 ( 60K) [CUDA0 ] ffn_up-19 ( 60K) [CUDA0 ]
node #720 ( MUL_MAT): ffn_out-19 ( 15K) [CUDA0 ]: blk.19.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-19 ( 60K) [CUDA0 ]
node #721 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-19 ( 15K) [CUDA0 ]
node #722 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.19.post_ffw_norm ( 15K) [CUDA0 ]
node #723 ( ADD): l_out-19 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-19 ( 15K) [CUDA0 ]
node #724 ( RMS_NORM): norm-20 ( 15K) [CUDA0 ]: l_out-19 ( 15K) [CUDA0 ]
node #725 ( MUL): attn_norm-20 ( 15K) [CUDA0 ]: norm-20 ( 15K) [CUDA0 ] blk.20.attn_norm.wei ( 15K) [CUDA0 ]
node #726 ( MUL_MAT): Qcur-20 ( 16K) [CUDA0 ]: blk.20.attn_q.weight ( 8M) [CUDA0 ] attn_norm-20 ( 15K) [CUDA0 ]
node #728 ( RMS_NORM): norm-20 ( 16K) [CUDA0 ]: Qcur-20 (reshaped) ( 16K) [CUDA0 ]
node #729 ( MUL): Qcur_normed-20 ( 16K) [CUDA0 ]: norm-20 ( 16K) [CUDA0 ] blk.20.attn_q_norm.w ( 1K) [CUDA0 ]
node #730 ( ROPE): Qcur-20 ( 16K) [CUDA0 ]: Qcur_normed-20 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #731 ( MUL_MAT): Kcur-20 ( 8K) [CUDA0 ]: blk.20.attn_k.weight ( 4M) [CUDA0 ] attn_norm-20 ( 15K) [CUDA0 ]
node #733 ( RMS_NORM): norm-20 ( 8K) [CUDA0 ]: Kcur-20 (reshaped) ( 8K) [CUDA0 ]
node #734 ( MUL): Kcur_normed-20 ( 8K) [CUDA0 ]: norm-20 ( 8K) [CUDA0 ] blk.20.attn_k_norm.w ( 1K) [CUDA0 ]
node #735 ( ROPE): Kcur-20 ( 8K) [CUDA0 ]: Kcur_normed-20 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #736 ( MUL_MAT): Vcur-20 ( 8K) [CUDA0 ]: blk.20.attn_v.weight ( 6M) [CUDA0 ] attn_norm-20 ( 15K) [CUDA0 ]
node #738 ( CPY): k_cache_view-20 (cop ( 2K) [CUDA0 ]: Kcur-20 ( 8K) [CUDA0 ] k_cache_view-20 ( 2K) [CUDA0 ]
node #740 ( CPY): v_cache_view-20 (cop ( 2K) [CUDA0 ]: Vcur-20 ( 8K) [CUDA0 ] v_cache_view-20 ( 2K) [CUDA0 ]

SPLIT #42: CPU # 3 inputs: [q-20 ( 16K)] [k-20 ( 544K)] [v-20 ( 544K)]

node #744 (FLASH_ATTN): node_744 ( 16K) [ CPU ]: CPU#q-20#0 ( 16K) [ NULL ] CPU#k-20#0 ( 544K) [ NULL ] CPU#v-20#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #43: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #746 ( MUL_MAT): kqv_out-20 ( 15K) [CUDA0 ]: blk.20.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #747 ( RMS_NORM): norm-20 ( 15K) [CUDA0 ]: kqv_out-20 ( 15K) [CUDA0 ]
node #748 ( MUL): attn_post_norm-20 ( 15K) [CUDA0 ]: norm-20 ( 15K) [CUDA0 ] blk.20.post_attentio ( 15K) [CUDA0 ]
node #749 ( ADD): sa_out-20 ( 15K) [CUDA0 ]: attn_post_norm-20 ( 15K) [CUDA0 ] l_out-19 ( 15K) [CUDA0 ]
node #750 ( RMS_NORM): norm-20 ( 15K) [CUDA0 ]: sa_out-20 ( 15K) [CUDA0 ]
node #751 ( MUL): ffn_norm-20 ( 15K) [CUDA0 ]: norm-20 ( 15K) [CUDA0 ] blk.20.ffn_norm.weig ( 15K) [CUDA0 ]
node #752 ( MUL_MAT): ffn_gate-20 ( 60K) [CUDA0 ]: blk.20.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-20 ( 15K) [CUDA0 ]
node #753 ( UNARY): ffn_gelu-20 ( 60K) [CUDA0 ]: ffn_gate-20 ( 60K) [CUDA0 ]
node #754 ( MUL_MAT): ffn_up-20 ( 60K) [CUDA0 ]: blk.20.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-20 ( 15K) [CUDA0 ]
node #755 ( MUL): ffn_gate_par-20 ( 60K) [CUDA0 ]: ffn_gelu-20 ( 60K) [CUDA0 ] ffn_up-20 ( 60K) [CUDA0 ]
node #756 ( MUL_MAT): ffn_out-20 ( 15K) [CUDA0 ]: blk.20.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-20 ( 60K) [CUDA0 ]
node #757 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-20 ( 15K) [CUDA0 ]
node #758 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.20.post_ffw_norm ( 15K) [CUDA0 ]
node #759 ( ADD): l_out-20 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-20 ( 15K) [CUDA0 ]
node #760 ( RMS_NORM): norm-21 ( 15K) [CUDA0 ]: l_out-20 ( 15K) [CUDA0 ]
node #761 ( MUL): attn_norm-21 ( 15K) [CUDA0 ]: norm-21 ( 15K) [CUDA0 ] blk.21.attn_norm.wei ( 15K) [CUDA0 ]
node #762 ( MUL_MAT): Qcur-21 ( 16K) [CUDA0 ]: blk.21.attn_q.weight ( 8M) [CUDA0 ] attn_norm-21 ( 15K) [CUDA0 ]
node #764 ( RMS_NORM): norm-21 ( 16K) [CUDA0 ]: Qcur-21 (reshaped) ( 16K) [CUDA0 ]
node #765 ( MUL): Qcur_normed-21 ( 16K) [CUDA0 ]: norm-21 ( 16K) [CUDA0 ] blk.21.attn_q_norm.w ( 1K) [CUDA0 ]
node #766 ( ROPE): Qcur-21 ( 16K) [CUDA0 ]: Qcur_normed-21 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #767 ( MUL_MAT): Kcur-21 ( 8K) [CUDA0 ]: blk.21.attn_k.weight ( 4M) [CUDA0 ] attn_norm-21 ( 15K) [CUDA0 ]
node #769 ( RMS_NORM): norm-21 ( 8K) [CUDA0 ]: Kcur-21 (reshaped) ( 8K) [CUDA0 ]
node #770 ( MUL): Kcur_normed-21 ( 8K) [CUDA0 ]: norm-21 ( 8K) [CUDA0 ] blk.21.attn_k_norm.w ( 1K) [CUDA0 ]
node #771 ( ROPE): Kcur-21 ( 8K) [CUDA0 ]: Kcur_normed-21 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #772 ( MUL_MAT): Vcur-21 ( 8K) [CUDA0 ]: blk.21.attn_v.weight ( 4M) [CUDA0 ] attn_norm-21 ( 15K) [CUDA0 ]
node #774 ( CPY): k_cache_view-21 (cop ( 2K) [CUDA0 ]: Kcur-21 ( 8K) [CUDA0 ] k_cache_view-21 ( 2K) [CUDA0 ]
node #776 ( CPY): v_cache_view-21 (cop ( 2K) [CUDA0 ]: Vcur-21 ( 8K) [CUDA0 ] v_cache_view-21 ( 2K) [CUDA0 ]

SPLIT #44: CPU # 3 inputs: [q-21 ( 16K)] [k-21 ( 544K)] [v-21 ( 544K)]

node #780 (FLASH_ATTN): node_780 ( 16K) [ CPU ]: CPU#q-21#0 ( 16K) [ NULL ] CPU#k-21#0 ( 544K) [ NULL ] CPU#v-21#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #45: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #782 ( MUL_MAT): kqv_out-21 ( 15K) [CUDA0 ]: blk.21.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #783 ( RMS_NORM): norm-21 ( 15K) [CUDA0 ]: kqv_out-21 ( 15K) [CUDA0 ]
node #784 ( MUL): attn_post_norm-21 ( 15K) [CUDA0 ]: norm-21 ( 15K) [CUDA0 ] blk.21.post_attentio ( 15K) [CUDA0 ]
node #785 ( ADD): sa_out-21 ( 15K) [CUDA0 ]: attn_post_norm-21 ( 15K) [CUDA0 ] l_out-20 ( 15K) [CUDA0 ]
node #786 ( RMS_NORM): norm-21 ( 15K) [CUDA0 ]: sa_out-21 ( 15K) [CUDA0 ]
node #787 ( MUL): ffn_norm-21 ( 15K) [CUDA0 ]: norm-21 ( 15K) [CUDA0 ] blk.21.ffn_norm.weig ( 15K) [CUDA0 ]
node #788 ( MUL_MAT): ffn_gate-21 ( 60K) [CUDA0 ]: blk.21.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-21 ( 15K) [CUDA0 ]
node #789 ( UNARY): ffn_gelu-21 ( 60K) [CUDA0 ]: ffn_gate-21 ( 60K) [CUDA0 ]
node #790 ( MUL_MAT): ffn_up-21 ( 60K) [CUDA0 ]: blk.21.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-21 ( 15K) [CUDA0 ]
node #791 ( MUL): ffn_gate_par-21 ( 60K) [CUDA0 ]: ffn_gelu-21 ( 60K) [CUDA0 ] ffn_up-21 ( 60K) [CUDA0 ]
node #792 ( MUL_MAT): ffn_out-21 ( 15K) [CUDA0 ]: blk.21.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-21 ( 60K) [CUDA0 ]
node #793 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-21 ( 15K) [CUDA0 ]
node #794 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.21.post_ffw_norm ( 15K) [CUDA0 ]
node #795 ( ADD): l_out-21 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-21 ( 15K) [CUDA0 ]
node #796 ( RMS_NORM): norm-22 ( 15K) [CUDA0 ]: l_out-21 ( 15K) [CUDA0 ]
node #797 ( MUL): attn_norm-22 ( 15K) [CUDA0 ]: norm-22 ( 15K) [CUDA0 ] blk.22.attn_norm.wei ( 15K) [CUDA0 ]
node #798 ( MUL_MAT): Qcur-22 ( 16K) [CUDA0 ]: blk.22.attn_q.weight ( 8M) [CUDA0 ] attn_norm-22 ( 15K) [CUDA0 ]
node #800 ( RMS_NORM): norm-22 ( 16K) [CUDA0 ]: Qcur-22 (reshaped) ( 16K) [CUDA0 ]
node #801 ( MUL): Qcur_normed-22 ( 16K) [CUDA0 ]: norm-22 ( 16K) [CUDA0 ] blk.22.attn_q_norm.w ( 1K) [CUDA0 ]
node #802 ( ROPE): Qcur-22 ( 16K) [CUDA0 ]: Qcur_normed-22 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #803 ( MUL_MAT): Kcur-22 ( 8K) [CUDA0 ]: blk.22.attn_k.weight ( 4M) [CUDA0 ] attn_norm-22 ( 15K) [CUDA0 ]
node #805 ( RMS_NORM): norm-22 ( 8K) [CUDA0 ]: Kcur-22 (reshaped) ( 8K) [CUDA0 ]
node #806 ( MUL): Kcur_normed-22 ( 8K) [CUDA0 ]: norm-22 ( 8K) [CUDA0 ] blk.22.attn_k_norm.w ( 1K) [CUDA0 ]
node #807 ( ROPE): Kcur-22 ( 8K) [CUDA0 ]: Kcur_normed-22 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #808 ( MUL_MAT): Vcur-22 ( 8K) [CUDA0 ]: blk.22.attn_v.weight ( 4M) [CUDA0 ] attn_norm-22 ( 15K) [CUDA0 ]
node #810 ( CPY): k_cache_view-22 (cop ( 2K) [CUDA0 ]: Kcur-22 ( 8K) [CUDA0 ] k_cache_view-22 ( 2K) [CUDA0 ]
node #812 ( CPY): v_cache_view-22 (cop ( 2K) [CUDA0 ]: Vcur-22 ( 8K) [CUDA0 ] v_cache_view-22 ( 2K) [CUDA0 ]

SPLIT #46: CPU # 3 inputs: [q-22 ( 16K)] [k-22 ( 544K)] [v-22 ( 544K)]

node #816 (FLASH_ATTN): node_816 ( 16K) [ CPU ]: CPU#q-22#0 ( 16K) [ NULL ] CPU#k-22#0 ( 544K) [ NULL ] CPU#v-22#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #47: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #818 ( MUL_MAT): kqv_out-22 ( 15K) [CUDA0 ]: blk.22.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #819 ( RMS_NORM): norm-22 ( 15K) [CUDA0 ]: kqv_out-22 ( 15K) [CUDA0 ]
node #820 ( MUL): attn_post_norm-22 ( 15K) [CUDA0 ]: norm-22 ( 15K) [CUDA0 ] blk.22.post_attentio ( 15K) [CUDA0 ]
node #821 ( ADD): sa_out-22 ( 15K) [CUDA0 ]: attn_post_norm-22 ( 15K) [CUDA0 ] l_out-21 ( 15K) [CUDA0 ]
node #822 ( RMS_NORM): norm-22 ( 15K) [CUDA0 ]: sa_out-22 ( 15K) [CUDA0 ]
node #823 ( MUL): ffn_norm-22 ( 15K) [CUDA0 ]: norm-22 ( 15K) [CUDA0 ] blk.22.ffn_norm.weig ( 15K) [CUDA0 ]
node #824 ( MUL_MAT): ffn_gate-22 ( 60K) [CUDA0 ]: blk.22.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-22 ( 15K) [CUDA0 ]
node #825 ( UNARY): ffn_gelu-22 ( 60K) [CUDA0 ]: ffn_gate-22 ( 60K) [CUDA0 ]
node #826 ( MUL_MAT): ffn_up-22 ( 60K) [CUDA0 ]: blk.22.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-22 ( 15K) [CUDA0 ]
node #827 ( MUL): ffn_gate_par-22 ( 60K) [CUDA0 ]: ffn_gelu-22 ( 60K) [CUDA0 ] ffn_up-22 ( 60K) [CUDA0 ]
node #828 ( MUL_MAT): ffn_out-22 ( 15K) [CUDA0 ]: blk.22.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-22 ( 60K) [CUDA0 ]
node #829 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-22 ( 15K) [CUDA0 ]
node #830 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.22.post_ffw_norm ( 15K) [CUDA0 ]
node #831 ( ADD): l_out-22 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-22 ( 15K) [CUDA0 ]
node #832 ( RMS_NORM): norm-23 ( 15K) [CUDA0 ]: l_out-22 ( 15K) [CUDA0 ]
node #833 ( MUL): attn_norm-23 ( 15K) [CUDA0 ]: norm-23 ( 15K) [CUDA0 ] blk.23.attn_norm.wei ( 15K) [CUDA0 ]
node #834 ( MUL_MAT): Qcur-23 ( 16K) [CUDA0 ]: blk.23.attn_q.weight ( 8M) [CUDA0 ] attn_norm-23 ( 15K) [CUDA0 ]
node #836 ( RMS_NORM): norm-23 ( 16K) [CUDA0 ]: Qcur-23 (reshaped) ( 16K) [CUDA0 ]
node #837 ( MUL): Qcur_normed-23 ( 16K) [CUDA0 ]: norm-23 ( 16K) [CUDA0 ] blk.23.attn_q_norm.w ( 1K) [CUDA0 ]
node #838 ( ROPE): Qcur-23 ( 16K) [CUDA0 ]: Qcur_normed-23 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #839 ( MUL_MAT): Kcur-23 ( 8K) [CUDA0 ]: blk.23.attn_k.weight ( 4M) [CUDA0 ] attn_norm-23 ( 15K) [CUDA0 ]
node #841 ( RMS_NORM): norm-23 ( 8K) [CUDA0 ]: Kcur-23 (reshaped) ( 8K) [CUDA0 ]
node #842 ( MUL): Kcur_normed-23 ( 8K) [CUDA0 ]: norm-23 ( 8K) [CUDA0 ] blk.23.attn_k_norm.w ( 1K) [CUDA0 ]
node #843 ( ROPE): Kcur-23 ( 8K) [CUDA0 ]: Kcur_normed-23 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #844 ( MUL_MAT): Vcur-23 ( 8K) [CUDA0 ]: blk.23.attn_v.weight ( 6M) [CUDA0 ] attn_norm-23 ( 15K) [CUDA0 ]
node #846 ( CPY): k_cache_view-23 (cop ( 2K) [CUDA0 ]: Kcur-23 ( 8K) [CUDA0 ] k_cache_view-23 ( 2K) [CUDA0 ]
node #848 ( CPY): v_cache_view-23 (cop ( 2K) [CUDA0 ]: Vcur-23 ( 8K) [CUDA0 ] v_cache_view-23 ( 2K) [CUDA0 ]

SPLIT #48: CPU # 3 inputs: [q-23 ( 16K)] [k-23 ( 544K)] [v-23 ( 544K)]

node #852 (FLASH_ATTN): node_852 ( 16K) [ CPU ]: CPU#q-23#0 ( 16K) [ NULL ] CPU#k-23#0 ( 544K) [ NULL ] CPU#v-23#0 ( 544K) [ NULL ] CPU#KQ_mask (copy)#0 ( 32K) [ NULL ]

SPLIT #49: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #854 ( MUL_MAT): kqv_out-23 ( 15K) [CUDA0 ]: blk.23.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #855 ( RMS_NORM): norm-23 ( 15K) [CUDA0 ]: kqv_out-23 ( 15K) [CUDA0 ]
node #856 ( MUL): attn_post_norm-23 ( 15K) [CUDA0 ]: norm-23 ( 15K) [CUDA0 ] blk.23.post_attentio ( 15K) [CUDA0 ]
node #857 ( ADD): sa_out-23 ( 15K) [CUDA0 ]: attn_post_norm-23 ( 15K) [CUDA0 ] l_out-22 ( 15K) [CUDA0 ]
node #858 ( RMS_NORM): norm-23 ( 15K) [CUDA0 ]: sa_out-23 ( 15K) [CUDA0 ]
node #859 ( MUL): ffn_norm-23 ( 15K) [CUDA0 ]: norm-23 ( 15K) [CUDA0 ] blk.23.ffn_norm.weig ( 15K) [CUDA0 ]
node #860 ( MUL_MAT): ffn_gate-23 ( 60K) [CUDA0 ]: blk.23.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-23 ( 15K) [CUDA0 ]
node #861 ( UNARY): ffn_gelu-23 ( 60K) [CUDA0 ]: ffn_gate-23 ( 60K) [CUDA0 ]
node #862 ( MUL_MAT): ffn_up-23 ( 60K) [CUDA0 ]: blk.23.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-23 ( 15K) [CUDA0 ]
node #863 ( MUL): ffn_gate_par-23 ( 60K) [CUDA0 ]: ffn_gelu-23 ( 60K) [CUDA0 ] ffn_up-23 ( 60K) [CUDA0 ]
node #864 ( MUL_MAT): ffn_out-23 ( 15K) [CUDA0 ]: blk.23.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-23 ( 60K) [CUDA0 ]
node #865 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-23 ( 15K) [CUDA0 ]
node #866 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.23.post_ffw_norm ( 15K) [CUDA0 ]
node #867 ( ADD): l_out-23 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-23 ( 15K) [CUDA0 ]
node #868 ( RMS_NORM): norm-24 ( 15K) [CUDA0 ]: l_out-23 ( 15K) [CUDA0 ]
node #869 ( MUL): attn_norm-24 ( 15K) [CUDA0 ]: norm-24 ( 15K) [CUDA0 ] blk.24.attn_norm.wei ( 15K) [CUDA0 ]
node #870 ( MUL_MAT): Qcur-24 ( 16K) [CUDA0 ]: blk.24.attn_q.weight ( 8M) [CUDA0 ] attn_norm-24 ( 15K) [CUDA0 ]
node #872 ( RMS_NORM): norm-24 ( 16K) [CUDA0 ]: Qcur-24 (reshaped) ( 16K) [CUDA0 ]
node #873 ( MUL): Qcur_normed-24 ( 16K) [CUDA0 ]: norm-24 ( 16K) [CUDA0 ] blk.24.attn_q_norm.w ( 1K) [CUDA0 ]
node #874 ( ROPE): Qcur-24 ( 16K) [CUDA0 ]: Qcur_normed-24 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #875 ( MUL_MAT): Kcur-24 ( 8K) [CUDA0 ]: blk.24.attn_k.weight ( 4M) [CUDA0 ] attn_norm-24 ( 15K) [CUDA0 ]
node #877 ( RMS_NORM): norm-24 ( 8K) [CUDA0 ]: Kcur-24 (reshaped) ( 8K) [CUDA0 ]
node #878 ( MUL): Kcur_normed-24 ( 8K) [CUDA0 ]: norm-24 ( 8K) [CUDA0 ] blk.24.attn_k_norm.w ( 1K) [CUDA0 ]
node #879 ( ROPE): Kcur-24 ( 8K) [CUDA0 ]: Kcur_normed-24 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #880 ( MUL_MAT): Vcur-24 ( 8K) [CUDA0 ]: blk.24.attn_v.weight ( 4M) [CUDA0 ] attn_norm-24 ( 15K) [CUDA0 ]
node #882 ( CPY): k_cache_view-24 (cop ( 2K) [CUDA0 ]: Kcur-24 ( 8K) [CUDA0 ] k_cache_view-24 ( 2K) [CUDA0 ]
node #884 ( CPY): v_cache_view-24 (cop ( 2K) [CUDA0 ]: Vcur-24 ( 8K) [CUDA0 ] v_cache_view-24 ( 2K) [CUDA0 ]

SPLIT #50: CPU # 3 inputs: [q-24 ( 16K)] [k-24 ( 544K)] [v-24 ( 544K)]

node #888 (FLASH_ATTN): node_888 ( 16K) [ CPU ]: CPU#q-24#0 ( 16K) [ NULL ] CPU#k-24#0 ( 544K) [ NULL ] CPU#v-24#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #51: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #890 ( MUL_MAT): kqv_out-24 ( 15K) [CUDA0 ]: blk.24.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #891 ( RMS_NORM): norm-24 ( 15K) [CUDA0 ]: kqv_out-24 ( 15K) [CUDA0 ]
node #892 ( MUL): attn_post_norm-24 ( 15K) [CUDA0 ]: norm-24 ( 15K) [CUDA0 ] blk.24.post_attentio ( 15K) [CUDA0 ]
node #893 ( ADD): sa_out-24 ( 15K) [CUDA0 ]: attn_post_norm-24 ( 15K) [CUDA0 ] l_out-23 ( 15K) [CUDA0 ]
node #894 ( RMS_NORM): norm-24 ( 15K) [CUDA0 ]: sa_out-24 ( 15K) [CUDA0 ]
node #895 ( MUL): ffn_norm-24 ( 15K) [CUDA0 ]: norm-24 ( 15K) [CUDA0 ] blk.24.ffn_norm.weig ( 15K) [CUDA0 ]
node #896 ( MUL_MAT): ffn_gate-24 ( 60K) [CUDA0 ]: blk.24.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-24 ( 15K) [CUDA0 ]
node #897 ( UNARY): ffn_gelu-24 ( 60K) [CUDA0 ]: ffn_gate-24 ( 60K) [CUDA0 ]
node #898 ( MUL_MAT): ffn_up-24 ( 60K) [CUDA0 ]: blk.24.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-24 ( 15K) [CUDA0 ]
node #899 ( MUL): ffn_gate_par-24 ( 60K) [CUDA0 ]: ffn_gelu-24 ( 60K) [CUDA0 ] ffn_up-24 ( 60K) [CUDA0 ]
node #900 ( MUL_MAT): ffn_out-24 ( 15K) [CUDA0 ]: blk.24.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-24 ( 60K) [CUDA0 ]
node #901 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-24 ( 15K) [CUDA0 ]
node #902 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.24.post_ffw_norm ( 15K) [CUDA0 ]
node #903 ( ADD): l_out-24 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-24 ( 15K) [CUDA0 ]
node #904 ( RMS_NORM): norm-25 ( 15K) [CUDA0 ]: l_out-24 ( 15K) [CUDA0 ]
node #905 ( MUL): attn_norm-25 ( 15K) [CUDA0 ]: norm-25 ( 15K) [CUDA0 ] blk.25.attn_norm.wei ( 15K) [CUDA0 ]
node #906 ( MUL_MAT): Qcur-25 ( 16K) [CUDA0 ]: blk.25.attn_q.weight ( 8M) [CUDA0 ] attn_norm-25 ( 15K) [CUDA0 ]
node #908 ( RMS_NORM): norm-25 ( 16K) [CUDA0 ]: Qcur-25 (reshaped) ( 16K) [CUDA0 ]
node #909 ( MUL): Qcur_normed-25 ( 16K) [CUDA0 ]: norm-25 ( 16K) [CUDA0 ] blk.25.attn_q_norm.w ( 1K) [CUDA0 ]
node #910 ( ROPE): Qcur-25 ( 16K) [CUDA0 ]: Qcur_normed-25 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #911 ( MUL_MAT): Kcur-25 ( 8K) [CUDA0 ]: blk.25.attn_k.weight ( 4M) [CUDA0 ] attn_norm-25 ( 15K) [CUDA0 ]
node #913 ( RMS_NORM): norm-25 ( 8K) [CUDA0 ]: Kcur-25 (reshaped) ( 8K) [CUDA0 ]
node #914 ( MUL): Kcur_normed-25 ( 8K) [CUDA0 ]: norm-25 ( 8K) [CUDA0 ] blk.25.attn_k_norm.w ( 1K) [CUDA0 ]
node #915 ( ROPE): Kcur-25 ( 8K) [CUDA0 ]: Kcur_normed-25 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #916 ( MUL_MAT): Vcur-25 ( 8K) [CUDA0 ]: blk.25.attn_v.weight ( 4M) [CUDA0 ] attn_norm-25 ( 15K) [CUDA0 ]
node #918 ( CPY): k_cache_view-25 (cop ( 2K) [CUDA0 ]: Kcur-25 ( 8K) [CUDA0 ] k_cache_view-25 ( 2K) [CUDA0 ]
node #920 ( CPY): v_cache_view-25 (cop ( 2K) [CUDA0 ]: Vcur-25 ( 8K) [CUDA0 ] v_cache_view-25 ( 2K) [CUDA0 ]

SPLIT #52: CPU # 3 inputs: [q-25 ( 16K)] [k-25 ( 544K)] [v-25 ( 544K)]

node #924 (FLASH_ATTN): node_924 ( 16K) [ CPU ]: CPU#q-25#0 ( 16K) [ NULL ] CPU#k-25#0 ( 544K) [ NULL ] CPU#v-25#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #53: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #926 ( MUL_MAT): kqv_out-25 ( 15K) [CUDA0 ]: blk.25.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #927 ( RMS_NORM): norm-25 ( 15K) [CUDA0 ]: kqv_out-25 ( 15K) [CUDA0 ]
node #928 ( MUL): attn_post_norm-25 ( 15K) [CUDA0 ]: norm-25 ( 15K) [CUDA0 ] blk.25.post_attentio ( 15K) [CUDA0 ]
node #929 ( ADD): sa_out-25 ( 15K) [CUDA0 ]: attn_post_norm-25 ( 15K) [CUDA0 ] l_out-24 ( 15K) [CUDA0 ]
node #930 ( RMS_NORM): norm-25 ( 15K) [CUDA0 ]: sa_out-25 ( 15K) [CUDA0 ]
node #931 ( MUL): ffn_norm-25 ( 15K) [CUDA0 ]: norm-25 ( 15K) [CUDA0 ] blk.25.ffn_norm.weig ( 15K) [CUDA0 ]
node #932 ( MUL_MAT): ffn_gate-25 ( 60K) [CUDA0 ]: blk.25.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-25 ( 15K) [CUDA0 ]
node #933 ( UNARY): ffn_gelu-25 ( 60K) [CUDA0 ]: ffn_gate-25 ( 60K) [CUDA0 ]
node #934 ( MUL_MAT): ffn_up-25 ( 60K) [CUDA0 ]: blk.25.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-25 ( 15K) [CUDA0 ]
node #935 ( MUL): ffn_gate_par-25 ( 60K) [CUDA0 ]: ffn_gelu-25 ( 60K) [CUDA0 ] ffn_up-25 ( 60K) [CUDA0 ]
node #936 ( MUL_MAT): ffn_out-25 ( 15K) [CUDA0 ]: blk.25.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-25 ( 60K) [CUDA0 ]
node #937 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-25 ( 15K) [CUDA0 ]
node #938 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.25.post_ffw_norm ( 15K) [CUDA0 ]
node #939 ( ADD): l_out-25 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-25 ( 15K) [CUDA0 ]
node #940 ( RMS_NORM): norm-26 ( 15K) [CUDA0 ]: l_out-25 ( 15K) [CUDA0 ]
node #941 ( MUL): attn_norm-26 ( 15K) [CUDA0 ]: norm-26 ( 15K) [CUDA0 ] blk.26.attn_norm.wei ( 15K) [CUDA0 ]
node #942 ( MUL_MAT): Qcur-26 ( 16K) [CUDA0 ]: blk.26.attn_q.weight ( 8M) [CUDA0 ] attn_norm-26 ( 15K) [CUDA0 ]
node #944 ( RMS_NORM): norm-26 ( 16K) [CUDA0 ]: Qcur-26 (reshaped) ( 16K) [CUDA0 ]
node #945 ( MUL): Qcur_normed-26 ( 16K) [CUDA0 ]: norm-26 ( 16K) [CUDA0 ] blk.26.attn_q_norm.w ( 1K) [CUDA0 ]
node #946 ( ROPE): Qcur-26 ( 16K) [CUDA0 ]: Qcur_normed-26 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #947 ( MUL_MAT): Kcur-26 ( 8K) [CUDA0 ]: blk.26.attn_k.weight ( 4M) [CUDA0 ] attn_norm-26 ( 15K) [CUDA0 ]
node #949 ( RMS_NORM): norm-26 ( 8K) [CUDA0 ]: Kcur-26 (reshaped) ( 8K) [CUDA0 ]
node #950 ( MUL): Kcur_normed-26 ( 8K) [CUDA0 ]: norm-26 ( 8K) [CUDA0 ] blk.26.attn_k_norm.w ( 1K) [CUDA0 ]
node #951 ( ROPE): Kcur-26 ( 8K) [CUDA0 ]: Kcur_normed-26 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #952 ( MUL_MAT): Vcur-26 ( 8K) [CUDA0 ]: blk.26.attn_v.weight ( 6M) [CUDA0 ] attn_norm-26 ( 15K) [CUDA0 ]
node #954 ( CPY): k_cache_view-26 (cop ( 2K) [CUDA0 ]: Kcur-26 ( 8K) [CUDA0 ] k_cache_view-26 ( 2K) [CUDA0 ]
node #956 ( CPY): v_cache_view-26 (cop ( 2K) [CUDA0 ]: Vcur-26 ( 8K) [CUDA0 ] v_cache_view-26 ( 2K) [CUDA0 ]

SPLIT #54: CPU # 3 inputs: [q-26 ( 16K)] [k-26 ( 544K)] [v-26 ( 544K)]

node #960 (FLASH_ATTN): node_960 ( 16K) [ CPU ]: CPU#q-26#0 ( 16K) [ NULL ] CPU#k-26#0 ( 544K) [ NULL ] CPU#v-26#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #55: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #962 ( MUL_MAT): kqv_out-26 ( 15K) [CUDA0 ]: blk.26.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #963 ( RMS_NORM): norm-26 ( 15K) [CUDA0 ]: kqv_out-26 ( 15K) [CUDA0 ]
node #964 ( MUL): attn_post_norm-26 ( 15K) [CUDA0 ]: norm-26 ( 15K) [CUDA0 ] blk.26.post_attentio ( 15K) [CUDA0 ]
node #965 ( ADD): sa_out-26 ( 15K) [CUDA0 ]: attn_post_norm-26 ( 15K) [CUDA0 ] l_out-25 ( 15K) [CUDA0 ]
node #966 ( RMS_NORM): norm-26 ( 15K) [CUDA0 ]: sa_out-26 ( 15K) [CUDA0 ]
node #967 ( MUL): ffn_norm-26 ( 15K) [CUDA0 ]: norm-26 ( 15K) [CUDA0 ] blk.26.ffn_norm.weig ( 15K) [CUDA0 ]
node #968 ( MUL_MAT): ffn_gate-26 ( 60K) [CUDA0 ]: blk.26.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-26 ( 15K) [CUDA0 ]
node #969 ( UNARY): ffn_gelu-26 ( 60K) [CUDA0 ]: ffn_gate-26 ( 60K) [CUDA0 ]
node #970 ( MUL_MAT): ffn_up-26 ( 60K) [CUDA0 ]: blk.26.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-26 ( 15K) [CUDA0 ]
node #971 ( MUL): ffn_gate_par-26 ( 60K) [CUDA0 ]: ffn_gelu-26 ( 60K) [CUDA0 ] ffn_up-26 ( 60K) [CUDA0 ]
node #972 ( MUL_MAT): ffn_out-26 ( 15K) [CUDA0 ]: blk.26.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-26 ( 60K) [CUDA0 ]
node #973 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-26 ( 15K) [CUDA0 ]
node #974 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.26.post_ffw_norm ( 15K) [CUDA0 ]
node #975 ( ADD): l_out-26 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-26 ( 15K) [CUDA0 ]
node #976 ( RMS_NORM): norm-27 ( 15K) [CUDA0 ]: l_out-26 ( 15K) [CUDA0 ]
node #977 ( MUL): attn_norm-27 ( 15K) [CUDA0 ]: norm-27 ( 15K) [CUDA0 ] blk.27.attn_norm.wei ( 15K) [CUDA0 ]
node #978 ( MUL_MAT): Qcur-27 ( 16K) [CUDA0 ]: blk.27.attn_q.weight ( 8M) [CUDA0 ] attn_norm-27 ( 15K) [CUDA0 ]
node #980 ( RMS_NORM): norm-27 ( 16K) [CUDA0 ]: Qcur-27 (reshaped) ( 16K) [CUDA0 ]
node #981 ( MUL): Qcur_normed-27 ( 16K) [CUDA0 ]: norm-27 ( 16K) [CUDA0 ] blk.27.attn_q_norm.w ( 1K) [CUDA0 ]
node #982 ( ROPE): Qcur-27 ( 16K) [CUDA0 ]: Qcur_normed-27 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #983 ( MUL_MAT): Kcur-27 ( 8K) [CUDA0 ]: blk.27.attn_k.weight ( 4M) [CUDA0 ] attn_norm-27 ( 15K) [CUDA0 ]
node #985 ( RMS_NORM): norm-27 ( 8K) [CUDA0 ]: Kcur-27 (reshaped) ( 8K) [CUDA0 ]
node #986 ( MUL): Kcur_normed-27 ( 8K) [CUDA0 ]: norm-27 ( 8K) [CUDA0 ] blk.27.attn_k_norm.w ( 1K) [CUDA0 ]
node #987 ( ROPE): Kcur-27 ( 8K) [CUDA0 ]: Kcur_normed-27 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #988 ( MUL_MAT): Vcur-27 ( 8K) [CUDA0 ]: blk.27.attn_v.weight ( 4M) [CUDA0 ] attn_norm-27 ( 15K) [CUDA0 ]
node #990 ( CPY): k_cache_view-27 (cop ( 2K) [CUDA0 ]: Kcur-27 ( 8K) [CUDA0 ] k_cache_view-27 ( 2K) [CUDA0 ]
node #992 ( CPY): v_cache_view-27 (cop ( 2K) [CUDA0 ]: Vcur-27 ( 8K) [CUDA0 ] v_cache_view-27 ( 2K) [CUDA0 ]

SPLIT #56: CPU # 3 inputs: [q-27 ( 16K)] [k-27 ( 544K)] [v-27 ( 544K)]

node #996 (FLASH_ATTN): node_996 ( 16K) [ CPU ]: CPU#q-27#0 ( 16K) [ NULL ] CPU#k-27#0 ( 544K) [ NULL ] CPU#v-27#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #57: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #998 ( MUL_MAT): kqv_out-27 ( 15K) [CUDA0 ]: blk.27.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #999 ( RMS_NORM): norm-27 ( 15K) [CUDA0 ]: kqv_out-27 ( 15K) [CUDA0 ]
node #1000 ( MUL): attn_post_norm-27 ( 15K) [CUDA0 ]: norm-27 ( 15K) [CUDA0 ] blk.27.post_attentio ( 15K) [CUDA0 ]
node #1001 ( ADD): sa_out-27 ( 15K) [CUDA0 ]: attn_post_norm-27 ( 15K) [CUDA0 ] l_out-26 ( 15K) [CUDA0 ]
node #1002 ( RMS_NORM): norm-27 ( 15K) [CUDA0 ]: sa_out-27 ( 15K) [CUDA0 ]
node #1003 ( MUL): ffn_norm-27 ( 15K) [CUDA0 ]: norm-27 ( 15K) [CUDA0 ] blk.27.ffn_norm.weig ( 15K) [CUDA0 ]
node #1004 ( MUL_MAT): ffn_gate-27 ( 60K) [CUDA0 ]: blk.27.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-27 ( 15K) [CUDA0 ]
node #1005 ( UNARY): ffn_gelu-27 ( 60K) [CUDA0 ]: ffn_gate-27 ( 60K) [CUDA0 ]
node #1006 ( MUL_MAT): ffn_up-27 ( 60K) [CUDA0 ]: blk.27.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-27 ( 15K) [CUDA0 ]
node #1007 ( MUL): ffn_gate_par-27 ( 60K) [CUDA0 ]: ffn_gelu-27 ( 60K) [CUDA0 ] ffn_up-27 ( 60K) [CUDA0 ]
node #1008 ( MUL_MAT): ffn_out-27 ( 15K) [CUDA0 ]: blk.27.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-27 ( 60K) [CUDA0 ]
node #1009 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-27 ( 15K) [CUDA0 ]
node #1010 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.27.post_ffw_norm ( 15K) [CUDA0 ]
node #1011 ( ADD): l_out-27 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-27 ( 15K) [CUDA0 ]
node #1012 ( RMS_NORM): norm-28 ( 15K) [CUDA0 ]: l_out-27 ( 15K) [CUDA0 ]
node #1013 ( MUL): attn_norm-28 ( 15K) [CUDA0 ]: norm-28 ( 15K) [CUDA0 ] blk.28.attn_norm.wei ( 15K) [CUDA0 ]
node #1014 ( MUL_MAT): Qcur-28 ( 16K) [CUDA0 ]: blk.28.attn_q.weight ( 8M) [CUDA0 ] attn_norm-28 ( 15K) [CUDA0 ]
node #1016 ( RMS_NORM): norm-28 ( 16K) [CUDA0 ]: Qcur-28 (reshaped) ( 16K) [CUDA0 ]
node #1017 ( MUL): Qcur_normed-28 ( 16K) [CUDA0 ]: norm-28 ( 16K) [CUDA0 ] blk.28.attn_q_norm.w ( 1K) [CUDA0 ]
node #1018 ( ROPE): Qcur-28 ( 16K) [CUDA0 ]: Qcur_normed-28 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1019 ( MUL_MAT): Kcur-28 ( 8K) [CUDA0 ]: blk.28.attn_k.weight ( 4M) [CUDA0 ] attn_norm-28 ( 15K) [CUDA0 ]
node #1021 ( RMS_NORM): norm-28 ( 8K) [CUDA0 ]: Kcur-28 (reshaped) ( 8K) [CUDA0 ]
node #1022 ( MUL): Kcur_normed-28 ( 8K) [CUDA0 ]: norm-28 ( 8K) [CUDA0 ] blk.28.attn_k_norm.w ( 1K) [CUDA0 ]
node #1023 ( ROPE): Kcur-28 ( 8K) [CUDA0 ]: Kcur_normed-28 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1024 ( MUL_MAT): Vcur-28 ( 8K) [CUDA0 ]: blk.28.attn_v.weight ( 4M) [CUDA0 ] attn_norm-28 ( 15K) [CUDA0 ]
node #1026 ( CPY): k_cache_view-28 (cop ( 2K) [CUDA0 ]: Kcur-28 ( 8K) [CUDA0 ] k_cache_view-28 ( 2K) [CUDA0 ]
node #1028 ( CPY): v_cache_view-28 (cop ( 2K) [CUDA0 ]: Vcur-28 ( 8K) [CUDA0 ] v_cache_view-28 ( 2K) [CUDA0 ]

SPLIT #58: CPU # 3 inputs: [q-28 ( 16K)] [k-28 ( 544K)] [v-28 ( 544K)]

node #1032 (FLASH_ATTN): node_1032 ( 16K) [ CPU ]: CPU#q-28#0 ( 16K) [ NULL ] CPU#k-28#0 ( 544K) [ NULL ] CPU#v-28#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #59: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1034 ( MUL_MAT): kqv_out-28 ( 15K) [CUDA0 ]: blk.28.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1035 ( RMS_NORM): norm-28 ( 15K) [CUDA0 ]: kqv_out-28 ( 15K) [CUDA0 ]
node #1036 ( MUL): attn_post_norm-28 ( 15K) [CUDA0 ]: norm-28 ( 15K) [CUDA0 ] blk.28.post_attentio ( 15K) [CUDA0 ]
node #1037 ( ADD): sa_out-28 ( 15K) [CUDA0 ]: attn_post_norm-28 ( 15K) [CUDA0 ] l_out-27 ( 15K) [CUDA0 ]
node #1038 ( RMS_NORM): norm-28 ( 15K) [CUDA0 ]: sa_out-28 ( 15K) [CUDA0 ]
node #1039 ( MUL): ffn_norm-28 ( 15K) [CUDA0 ]: norm-28 ( 15K) [CUDA0 ] blk.28.ffn_norm.weig ( 15K) [CUDA0 ]
node #1040 ( MUL_MAT): ffn_gate-28 ( 60K) [CUDA0 ]: blk.28.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-28 ( 15K) [CUDA0 ]
node #1041 ( UNARY): ffn_gelu-28 ( 60K) [CUDA0 ]: ffn_gate-28 ( 60K) [CUDA0 ]
node #1042 ( MUL_MAT): ffn_up-28 ( 60K) [CUDA0 ]: blk.28.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-28 ( 15K) [CUDA0 ]
node #1043 ( MUL): ffn_gate_par-28 ( 60K) [CUDA0 ]: ffn_gelu-28 ( 60K) [CUDA0 ] ffn_up-28 ( 60K) [CUDA0 ]
node #1044 ( MUL_MAT): ffn_out-28 ( 15K) [CUDA0 ]: blk.28.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-28 ( 60K) [CUDA0 ]
node #1045 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-28 ( 15K) [CUDA0 ]
node #1046 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.28.post_ffw_norm ( 15K) [CUDA0 ]
node #1047 ( ADD): l_out-28 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-28 ( 15K) [CUDA0 ]
node #1048 ( RMS_NORM): norm-29 ( 15K) [CUDA0 ]: l_out-28 ( 15K) [CUDA0 ]
node #1049 ( MUL): attn_norm-29 ( 15K) [CUDA0 ]: norm-29 ( 15K) [CUDA0 ] blk.29.attn_norm.wei ( 15K) [CUDA0 ]
node #1050 ( MUL_MAT): Qcur-29 ( 16K) [CUDA0 ]: blk.29.attn_q.weight ( 8M) [CUDA0 ] attn_norm-29 ( 15K) [CUDA0 ]
node #1052 ( RMS_NORM): norm-29 ( 16K) [CUDA0 ]: Qcur-29 (reshaped) ( 16K) [CUDA0 ]
node #1053 ( MUL): Qcur_normed-29 ( 16K) [CUDA0 ]: norm-29 ( 16K) [CUDA0 ] blk.29.attn_q_norm.w ( 1K) [CUDA0 ]
node #1054 ( ROPE): Qcur-29 ( 16K) [CUDA0 ]: Qcur_normed-29 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1055 ( MUL_MAT): Kcur-29 ( 8K) [CUDA0 ]: blk.29.attn_k.weight ( 4M) [CUDA0 ] attn_norm-29 ( 15K) [CUDA0 ]
node #1057 ( RMS_NORM): norm-29 ( 8K) [CUDA0 ]: Kcur-29 (reshaped) ( 8K) [CUDA0 ]
node #1058 ( MUL): Kcur_normed-29 ( 8K) [CUDA0 ]: norm-29 ( 8K) [CUDA0 ] blk.29.attn_k_norm.w ( 1K) [CUDA0 ]
node #1059 ( ROPE): Kcur-29 ( 8K) [CUDA0 ]: Kcur_normed-29 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1060 ( MUL_MAT): Vcur-29 ( 8K) [CUDA0 ]: blk.29.attn_v.weight ( 6M) [CUDA0 ] attn_norm-29 ( 15K) [CUDA0 ]
node #1062 ( CPY): k_cache_view-29 (cop ( 2K) [CUDA0 ]: Kcur-29 ( 8K) [CUDA0 ] k_cache_view-29 ( 2K) [CUDA0 ]
node #1064 ( CPY): v_cache_view-29 (cop ( 2K) [CUDA0 ]: Vcur-29 ( 8K) [CUDA0 ] v_cache_view-29 ( 2K) [CUDA0 ]

SPLIT #60: CPU # 3 inputs: [q-29 ( 16K)] [k-29 ( 544K)] [v-29 ( 544K)]

node #1068 (FLASH_ATTN): node_1068 ( 16K) [ CPU ]: CPU#q-29#0 ( 16K) [ NULL ] CPU#k-29#0 ( 544K) [ NULL ] CPU#v-29#0 ( 544K) [ NULL ] CPU#KQ_mask (copy)#0 ( 32K) [ NULL ]

SPLIT #61: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1070 ( MUL_MAT): kqv_out-29 ( 15K) [CUDA0 ]: blk.29.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1071 ( RMS_NORM): norm-29 ( 15K) [CUDA0 ]: kqv_out-29 ( 15K) [CUDA0 ]
node #1072 ( MUL): attn_post_norm-29 ( 15K) [CUDA0 ]: norm-29 ( 15K) [CUDA0 ] blk.29.post_attentio ( 15K) [CUDA0 ]
node #1073 ( ADD): sa_out-29 ( 15K) [CUDA0 ]: attn_post_norm-29 ( 15K) [CUDA0 ] l_out-28 ( 15K) [CUDA0 ]
node #1074 ( RMS_NORM): norm-29 ( 15K) [CUDA0 ]: sa_out-29 ( 15K) [CUDA0 ]
node #1075 ( MUL): ffn_norm-29 ( 15K) [CUDA0 ]: norm-29 ( 15K) [CUDA0 ] blk.29.ffn_norm.weig ( 15K) [CUDA0 ]
node #1076 ( MUL_MAT): ffn_gate-29 ( 60K) [CUDA0 ]: blk.29.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-29 ( 15K) [CUDA0 ]
node #1077 ( UNARY): ffn_gelu-29 ( 60K) [CUDA0 ]: ffn_gate-29 ( 60K) [CUDA0 ]
node #1078 ( MUL_MAT): ffn_up-29 ( 60K) [CUDA0 ]: blk.29.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-29 ( 15K) [CUDA0 ]
node #1079 ( MUL): ffn_gate_par-29 ( 60K) [CUDA0 ]: ffn_gelu-29 ( 60K) [CUDA0 ] ffn_up-29 ( 60K) [CUDA0 ]
node #1080 ( MUL_MAT): ffn_out-29 ( 15K) [CUDA0 ]: blk.29.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-29 ( 60K) [CUDA0 ]
node #1081 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-29 ( 15K) [CUDA0 ]
node #1082 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.29.post_ffw_norm ( 15K) [CUDA0 ]
node #1083 ( ADD): l_out-29 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-29 ( 15K) [CUDA0 ]
node #1084 ( RMS_NORM): norm-30 ( 15K) [CUDA0 ]: l_out-29 ( 15K) [CUDA0 ]
node #1085 ( MUL): attn_norm-30 ( 15K) [CUDA0 ]: norm-30 ( 15K) [CUDA0 ] blk.30.attn_norm.wei ( 15K) [CUDA0 ]
node #1086 ( MUL_MAT): Qcur-30 ( 16K) [CUDA0 ]: blk.30.attn_q.weight ( 8M) [CUDA0 ] attn_norm-30 ( 15K) [CUDA0 ]
node #1088 ( RMS_NORM): norm-30 ( 16K) [CUDA0 ]: Qcur-30 (reshaped) ( 16K) [CUDA0 ]
node #1089 ( MUL): Qcur_normed-30 ( 16K) [CUDA0 ]: norm-30 ( 16K) [CUDA0 ] blk.30.attn_q_norm.w ( 1K) [CUDA0 ]
node #1090 ( ROPE): Qcur-30 ( 16K) [CUDA0 ]: Qcur_normed-30 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1091 ( MUL_MAT): Kcur-30 ( 8K) [CUDA0 ]: blk.30.attn_k.weight ( 4M) [CUDA0 ] attn_norm-30 ( 15K) [CUDA0 ]
node #1093 ( RMS_NORM): norm-30 ( 8K) [CUDA0 ]: Kcur-30 (reshaped) ( 8K) [CUDA0 ]
node #1094 ( MUL): Kcur_normed-30 ( 8K) [CUDA0 ]: norm-30 ( 8K) [CUDA0 ] blk.30.attn_k_norm.w ( 1K) [CUDA0 ]
node #1095 ( ROPE): Kcur-30 ( 8K) [CUDA0 ]: Kcur_normed-30 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1096 ( MUL_MAT): Vcur-30 ( 8K) [CUDA0 ]: blk.30.attn_v.weight ( 4M) [CUDA0 ] attn_norm-30 ( 15K) [CUDA0 ]
node #1098 ( CPY): k_cache_view-30 (cop ( 2K) [CUDA0 ]: Kcur-30 ( 8K) [CUDA0 ] k_cache_view-30 ( 2K) [CUDA0 ]
node #1100 ( CPY): v_cache_view-30 (cop ( 2K) [CUDA0 ]: Vcur-30 ( 8K) [CUDA0 ] v_cache_view-30 ( 2K) [CUDA0 ]

SPLIT #62: CPU # 3 inputs: [q-30 ( 16K)] [k-30 ( 544K)] [v-30 ( 544K)]

node #1104 (FLASH_ATTN): node_1104 ( 16K) [ CPU ]: CPU#q-30#0 ( 16K) [ NULL ] CPU#k-30#0 ( 544K) [ NULL ] CPU#v-30#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #63: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1106 ( MUL_MAT): kqv_out-30 ( 15K) [CUDA0 ]: blk.30.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1107 ( RMS_NORM): norm-30 ( 15K) [CUDA0 ]: kqv_out-30 ( 15K) [CUDA0 ]
node #1108 ( MUL): attn_post_norm-30 ( 15K) [CUDA0 ]: norm-30 ( 15K) [CUDA0 ] blk.30.post_attentio ( 15K) [CUDA0 ]
node #1109 ( ADD): sa_out-30 ( 15K) [CUDA0 ]: attn_post_norm-30 ( 15K) [CUDA0 ] l_out-29 ( 15K) [CUDA0 ]
node #1110 ( RMS_NORM): norm-30 ( 15K) [CUDA0 ]: sa_out-30 ( 15K) [CUDA0 ]
node #1111 ( MUL): ffn_norm-30 ( 15K) [CUDA0 ]: norm-30 ( 15K) [CUDA0 ] blk.30.ffn_norm.weig ( 15K) [CUDA0 ]
node #1112 ( MUL_MAT): ffn_gate-30 ( 60K) [CUDA0 ]: blk.30.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-30 ( 15K) [CUDA0 ]
node #1113 ( UNARY): ffn_gelu-30 ( 60K) [CUDA0 ]: ffn_gate-30 ( 60K) [CUDA0 ]
node #1114 ( MUL_MAT): ffn_up-30 ( 60K) [CUDA0 ]: blk.30.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-30 ( 15K) [CUDA0 ]
node #1115 ( MUL): ffn_gate_par-30 ( 60K) [CUDA0 ]: ffn_gelu-30 ( 60K) [CUDA0 ] ffn_up-30 ( 60K) [CUDA0 ]
node #1116 ( MUL_MAT): ffn_out-30 ( 15K) [CUDA0 ]: blk.30.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-30 ( 60K) [CUDA0 ]
node #1117 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-30 ( 15K) [CUDA0 ]
node #1118 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.30.post_ffw_norm ( 15K) [CUDA0 ]
node #1119 ( ADD): l_out-30 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-30 ( 15K) [CUDA0 ]
node #1120 ( RMS_NORM): norm-31 ( 15K) [CUDA0 ]: l_out-30 ( 15K) [CUDA0 ]
node #1121 ( MUL): attn_norm-31 ( 15K) [CUDA0 ]: norm-31 ( 15K) [CUDA0 ] blk.31.attn_norm.wei ( 15K) [CUDA0 ]
node #1122 ( MUL_MAT): Qcur-31 ( 16K) [CUDA0 ]: blk.31.attn_q.weight ( 8M) [CUDA0 ] attn_norm-31 ( 15K) [CUDA0 ]
node #1124 ( RMS_NORM): norm-31 ( 16K) [CUDA0 ]: Qcur-31 (reshaped) ( 16K) [CUDA0 ]
node #1125 ( MUL): Qcur_normed-31 ( 16K) [CUDA0 ]: norm-31 ( 16K) [CUDA0 ] blk.31.attn_q_norm.w ( 1K) [CUDA0 ]
node #1126 ( ROPE): Qcur-31 ( 16K) [CUDA0 ]: Qcur_normed-31 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1127 ( MUL_MAT): Kcur-31 ( 8K) [CUDA0 ]: blk.31.attn_k.weight ( 4M) [CUDA0 ] attn_norm-31 ( 15K) [CUDA0 ]
node #1129 ( RMS_NORM): norm-31 ( 8K) [CUDA0 ]: Kcur-31 (reshaped) ( 8K) [CUDA0 ]
node #1130 ( MUL): Kcur_normed-31 ( 8K) [CUDA0 ]: norm-31 ( 8K) [CUDA0 ] blk.31.attn_k_norm.w ( 1K) [CUDA0 ]
node #1131 ( ROPE): Kcur-31 ( 8K) [CUDA0 ]: Kcur_normed-31 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1132 ( MUL_MAT): Vcur-31 ( 8K) [CUDA0 ]: blk.31.attn_v.weight ( 4M) [CUDA0 ] attn_norm-31 ( 15K) [CUDA0 ]
node #1134 ( CPY): k_cache_view-31 (cop ( 2K) [CUDA0 ]: Kcur-31 ( 8K) [CUDA0 ] k_cache_view-31 ( 2K) [CUDA0 ]
node #1136 ( CPY): v_cache_view-31 (cop ( 2K) [CUDA0 ]: Vcur-31 ( 8K) [CUDA0 ] v_cache_view-31 ( 2K) [CUDA0 ]

SPLIT #64: CPU # 3 inputs: [q-31 ( 16K)] [k-31 ( 544K)] [v-31 ( 544K)]

node #1140 (FLASH_ATTN): node_1140 ( 16K) [ CPU ]: CPU#q-31#0 ( 16K) [ NULL ] CPU#k-31#0 ( 544K) [ NULL ] CPU#v-31#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #65: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1142 ( MUL_MAT): kqv_out-31 ( 15K) [CUDA0 ]: blk.31.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1143 ( RMS_NORM): norm-31 ( 15K) [CUDA0 ]: kqv_out-31 ( 15K) [CUDA0 ]
node #1144 ( MUL): attn_post_norm-31 ( 15K) [CUDA0 ]: norm-31 ( 15K) [CUDA0 ] blk.31.post_attentio ( 15K) [CUDA0 ]
node #1145 ( ADD): sa_out-31 ( 15K) [CUDA0 ]: attn_post_norm-31 ( 15K) [CUDA0 ] l_out-30 ( 15K) [CUDA0 ]
node #1146 ( RMS_NORM): norm-31 ( 15K) [CUDA0 ]: sa_out-31 ( 15K) [CUDA0 ]
node #1147 ( MUL): ffn_norm-31 ( 15K) [CUDA0 ]: norm-31 ( 15K) [CUDA0 ] blk.31.ffn_norm.weig ( 15K) [CUDA0 ]
node #1148 ( MUL_MAT): ffn_gate-31 ( 60K) [CUDA0 ]: blk.31.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-31 ( 15K) [CUDA0 ]
node #1149 ( UNARY): ffn_gelu-31 ( 60K) [CUDA0 ]: ffn_gate-31 ( 60K) [CUDA0 ]
node #1150 ( MUL_MAT): ffn_up-31 ( 60K) [CUDA0 ]: blk.31.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-31 ( 15K) [CUDA0 ]
node #1151 ( MUL): ffn_gate_par-31 ( 60K) [CUDA0 ]: ffn_gelu-31 ( 60K) [CUDA0 ] ffn_up-31 ( 60K) [CUDA0 ]
node #1152 ( MUL_MAT): ffn_out-31 ( 15K) [CUDA0 ]: blk.31.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-31 ( 60K) [CUDA0 ]
node #1153 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-31 ( 15K) [CUDA0 ]
node #1154 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.31.post_ffw_norm ( 15K) [CUDA0 ]
node #1155 ( ADD): l_out-31 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-31 ( 15K) [CUDA0 ]
node #1156 ( RMS_NORM): norm-32 ( 15K) [CUDA0 ]: l_out-31 ( 15K) [CUDA0 ]
node #1157 ( MUL): attn_norm-32 ( 15K) [CUDA0 ]: norm-32 ( 15K) [CUDA0 ] blk.32.attn_norm.wei ( 15K) [CUDA0 ]
node #1158 ( MUL_MAT): Qcur-32 ( 16K) [CUDA0 ]: blk.32.attn_q.weight ( 8M) [CUDA0 ] attn_norm-32 ( 15K) [CUDA0 ]
node #1160 ( RMS_NORM): norm-32 ( 16K) [CUDA0 ]: Qcur-32 (reshaped) ( 16K) [CUDA0 ]
node #1161 ( MUL): Qcur_normed-32 ( 16K) [CUDA0 ]: norm-32 ( 16K) [CUDA0 ] blk.32.attn_q_norm.w ( 1K) [CUDA0 ]
node #1162 ( ROPE): Qcur-32 ( 16K) [CUDA0 ]: Qcur_normed-32 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1163 ( MUL_MAT): Kcur-32 ( 8K) [CUDA0 ]: blk.32.attn_k.weight ( 4M) [CUDA0 ] attn_norm-32 ( 15K) [CUDA0 ]
node #1165 ( RMS_NORM): norm-32 ( 8K) [CUDA0 ]: Kcur-32 (reshaped) ( 8K) [CUDA0 ]
node #1166 ( MUL): Kcur_normed-32 ( 8K) [CUDA0 ]: norm-32 ( 8K) [CUDA0 ] blk.32.attn_k_norm.w ( 1K) [CUDA0 ]
node #1167 ( ROPE): Kcur-32 ( 8K) [CUDA0 ]: Kcur_normed-32 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1168 ( MUL_MAT): Vcur-32 ( 8K) [CUDA0 ]: blk.32.attn_v.weight ( 6M) [CUDA0 ] attn_norm-32 ( 15K) [CUDA0 ]
node #1170 ( CPY): k_cache_view-32 (cop ( 2K) [CUDA0 ]: Kcur-32 ( 8K) [CUDA0 ] k_cache_view-32 ( 2K) [CUDA0 ]
node #1172 ( CPY): v_cache_view-32 (cop ( 2K) [CUDA0 ]: Vcur-32 ( 8K) [CUDA0 ] v_cache_view-32 ( 2K) [CUDA0 ]

SPLIT #66: CPU # 3 inputs: [q-32 ( 16K)] [k-32 ( 544K)] [v-32 ( 544K)]

node #1176 (FLASH_ATTN): node_1176 ( 16K) [ CPU ]: CPU#q-32#0 ( 16K) [ NULL ] CPU#k-32#0 ( 544K) [ NULL ] CPU#v-32#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #67: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1178 ( MUL_MAT): kqv_out-32 ( 15K) [CUDA0 ]: blk.32.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1179 ( RMS_NORM): norm-32 ( 15K) [CUDA0 ]: kqv_out-32 ( 15K) [CUDA0 ]
node #1180 ( MUL): attn_post_norm-32 ( 15K) [CUDA0 ]: norm-32 ( 15K) [CUDA0 ] blk.32.post_attentio ( 15K) [CUDA0 ]
node #1181 ( ADD): sa_out-32 ( 15K) [CUDA0 ]: attn_post_norm-32 ( 15K) [CUDA0 ] l_out-31 ( 15K) [CUDA0 ]
node #1182 ( RMS_NORM): norm-32 ( 15K) [CUDA0 ]: sa_out-32 ( 15K) [CUDA0 ]
node #1183 ( MUL): ffn_norm-32 ( 15K) [CUDA0 ]: norm-32 ( 15K) [CUDA0 ] blk.32.ffn_norm.weig ( 15K) [CUDA0 ]
node #1184 ( MUL_MAT): ffn_gate-32 ( 60K) [CUDA0 ]: blk.32.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-32 ( 15K) [CUDA0 ]
node #1185 ( UNARY): ffn_gelu-32 ( 60K) [CUDA0 ]: ffn_gate-32 ( 60K) [CUDA0 ]
node #1186 ( MUL_MAT): ffn_up-32 ( 60K) [CUDA0 ]: blk.32.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-32 ( 15K) [CUDA0 ]
node #1187 ( MUL): ffn_gate_par-32 ( 60K) [CUDA0 ]: ffn_gelu-32 ( 60K) [CUDA0 ] ffn_up-32 ( 60K) [CUDA0 ]
node #1188 ( MUL_MAT): ffn_out-32 ( 15K) [CUDA0 ]: blk.32.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-32 ( 60K) [CUDA0 ]
node #1189 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-32 ( 15K) [CUDA0 ]
node #1190 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.32.post_ffw_norm ( 15K) [CUDA0 ]
node #1191 ( ADD): l_out-32 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-32 ( 15K) [CUDA0 ]
node #1192 ( RMS_NORM): norm-33 ( 15K) [CUDA0 ]: l_out-32 ( 15K) [CUDA0 ]
node #1193 ( MUL): attn_norm-33 ( 15K) [CUDA0 ]: norm-33 ( 15K) [CUDA0 ] blk.33.attn_norm.wei ( 15K) [CUDA0 ]
node #1194 ( MUL_MAT): Qcur-33 ( 16K) [CUDA0 ]: blk.33.attn_q.weight ( 8M) [CUDA0 ] attn_norm-33 ( 15K) [CUDA0 ]
node #1196 ( RMS_NORM): norm-33 ( 16K) [CUDA0 ]: Qcur-33 (reshaped) ( 16K) [CUDA0 ]
node #1197 ( MUL): Qcur_normed-33 ( 16K) [CUDA0 ]: norm-33 ( 16K) [CUDA0 ] blk.33.attn_q_norm.w ( 1K) [CUDA0 ]
node #1198 ( ROPE): Qcur-33 ( 16K) [CUDA0 ]: Qcur_normed-33 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1199 ( MUL_MAT): Kcur-33 ( 8K) [CUDA0 ]: blk.33.attn_k.weight ( 4M) [CUDA0 ] attn_norm-33 ( 15K) [CUDA0 ]
node #1201 ( RMS_NORM): norm-33 ( 8K) [CUDA0 ]: Kcur-33 (reshaped) ( 8K) [CUDA0 ]
node #1202 ( MUL): Kcur_normed-33 ( 8K) [CUDA0 ]: norm-33 ( 8K) [CUDA0 ] blk.33.attn_k_norm.w ( 1K) [CUDA0 ]
node #1203 ( ROPE): Kcur-33 ( 8K) [CUDA0 ]: Kcur_normed-33 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1204 ( MUL_MAT): Vcur-33 ( 8K) [CUDA0 ]: blk.33.attn_v.weight ( 4M) [CUDA0 ] attn_norm-33 ( 15K) [CUDA0 ]
node #1206 ( CPY): k_cache_view-33 (cop ( 2K) [CUDA0 ]: Kcur-33 ( 8K) [CUDA0 ] k_cache_view-33 ( 2K) [CUDA0 ]
node #1208 ( CPY): v_cache_view-33 (cop ( 2K) [CUDA0 ]: Vcur-33 ( 8K) [CUDA0 ] v_cache_view-33 ( 2K) [CUDA0 ]

SPLIT #68: CPU # 3 inputs: [q-33 ( 16K)] [k-33 ( 544K)] [v-33 ( 544K)]

node #1212 (FLASH_ATTN): node_1212 ( 16K) [ CPU ]: CPU#q-33#0 ( 16K) [ NULL ] CPU#k-33#0 ( 544K) [ NULL ] CPU#v-33#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #69: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1214 ( MUL_MAT): kqv_out-33 ( 15K) [CUDA0 ]: blk.33.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1215 ( RMS_NORM): norm-33 ( 15K) [CUDA0 ]: kqv_out-33 ( 15K) [CUDA0 ]
node #1216 ( MUL): attn_post_norm-33 ( 15K) [CUDA0 ]: norm-33 ( 15K) [CUDA0 ] blk.33.post_attentio ( 15K) [CUDA0 ]
node #1217 ( ADD): sa_out-33 ( 15K) [CUDA0 ]: attn_post_norm-33 ( 15K) [CUDA0 ] l_out-32 ( 15K) [CUDA0 ]
node #1218 ( RMS_NORM): norm-33 ( 15K) [CUDA0 ]: sa_out-33 ( 15K) [CUDA0 ]
node #1219 ( MUL): ffn_norm-33 ( 15K) [CUDA0 ]: norm-33 ( 15K) [CUDA0 ] blk.33.ffn_norm.weig ( 15K) [CUDA0 ]
node #1220 ( MUL_MAT): ffn_gate-33 ( 60K) [CUDA0 ]: blk.33.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-33 ( 15K) [CUDA0 ]
node #1221 ( UNARY): ffn_gelu-33 ( 60K) [CUDA0 ]: ffn_gate-33 ( 60K) [CUDA0 ]
node #1222 ( MUL_MAT): ffn_up-33 ( 60K) [CUDA0 ]: blk.33.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-33 ( 15K) [CUDA0 ]
node #1223 ( MUL): ffn_gate_par-33 ( 60K) [CUDA0 ]: ffn_gelu-33 ( 60K) [CUDA0 ] ffn_up-33 ( 60K) [CUDA0 ]
node #1224 ( MUL_MAT): ffn_out-33 ( 15K) [CUDA0 ]: blk.33.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-33 ( 60K) [CUDA0 ]
node #1225 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-33 ( 15K) [CUDA0 ]
node #1226 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.33.post_ffw_norm ( 15K) [CUDA0 ]
node #1227 ( ADD): l_out-33 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-33 ( 15K) [CUDA0 ]
node #1228 ( RMS_NORM): norm-34 ( 15K) [CUDA0 ]: l_out-33 ( 15K) [CUDA0 ]
node #1229 ( MUL): attn_norm-34 ( 15K) [CUDA0 ]: norm-34 ( 15K) [CUDA0 ] blk.34.attn_norm.wei ( 15K) [CUDA0 ]
node #1230 ( MUL_MAT): Qcur-34 ( 16K) [CUDA0 ]: blk.34.attn_q.weight ( 8M) [CUDA0 ] attn_norm-34 ( 15K) [CUDA0 ]
node #1232 ( RMS_NORM): norm-34 ( 16K) [CUDA0 ]: Qcur-34 (reshaped) ( 16K) [CUDA0 ]
node #1233 ( MUL): Qcur_normed-34 ( 16K) [CUDA0 ]: norm-34 ( 16K) [CUDA0 ] blk.34.attn_q_norm.w ( 1K) [CUDA0 ]
node #1234 ( ROPE): Qcur-34 ( 16K) [CUDA0 ]: Qcur_normed-34 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1235 ( MUL_MAT): Kcur-34 ( 8K) [CUDA0 ]: blk.34.attn_k.weight ( 4M) [CUDA0 ] attn_norm-34 ( 15K) [CUDA0 ]
node #1237 ( RMS_NORM): norm-34 ( 8K) [CUDA0 ]: Kcur-34 (reshaped) ( 8K) [CUDA0 ]
node #1238 ( MUL): Kcur_normed-34 ( 8K) [CUDA0 ]: norm-34 ( 8K) [CUDA0 ] blk.34.attn_k_norm.w ( 1K) [CUDA0 ]
node #1239 ( ROPE): Kcur-34 ( 8K) [CUDA0 ]: Kcur_normed-34 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1240 ( MUL_MAT): Vcur-34 ( 8K) [CUDA0 ]: blk.34.attn_v.weight ( 4M) [CUDA0 ] attn_norm-34 ( 15K) [CUDA0 ]
node #1242 ( CPY): k_cache_view-34 (cop ( 2K) [CUDA0 ]: Kcur-34 ( 8K) [CUDA0 ] k_cache_view-34 ( 2K) [CUDA0 ]
node #1244 ( CPY): v_cache_view-34 (cop ( 2K) [CUDA0 ]: Vcur-34 ( 8K) [CUDA0 ] v_cache_view-34 ( 2K) [CUDA0 ]

SPLIT #70: CPU # 3 inputs: [q-34 ( 16K)] [k-34 ( 544K)] [v-34 ( 544K)]

node #1248 (FLASH_ATTN): node_1248 ( 16K) [ CPU ]: CPU#q-34#0 ( 16K) [ NULL ] CPU#k-34#0 ( 544K) [ NULL ] CPU#v-34#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #71: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1250 ( MUL_MAT): kqv_out-34 ( 15K) [CUDA0 ]: blk.34.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1251 ( RMS_NORM): norm-34 ( 15K) [CUDA0 ]: kqv_out-34 ( 15K) [CUDA0 ]
node #1252 ( MUL): attn_post_norm-34 ( 15K) [CUDA0 ]: norm-34 ( 15K) [CUDA0 ] blk.34.post_attentio ( 15K) [CUDA0 ]
node #1253 ( ADD): sa_out-34 ( 15K) [CUDA0 ]: attn_post_norm-34 ( 15K) [CUDA0 ] l_out-33 ( 15K) [CUDA0 ]
node #1254 ( RMS_NORM): norm-34 ( 15K) [CUDA0 ]: sa_out-34 ( 15K) [CUDA0 ]
node #1255 ( MUL): ffn_norm-34 ( 15K) [CUDA0 ]: norm-34 ( 15K) [CUDA0 ] blk.34.ffn_norm.weig ( 15K) [CUDA0 ]
node #1256 ( MUL_MAT): ffn_gate-34 ( 60K) [CUDA0 ]: blk.34.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-34 ( 15K) [CUDA0 ]
node #1257 ( UNARY): ffn_gelu-34 ( 60K) [CUDA0 ]: ffn_gate-34 ( 60K) [CUDA0 ]
node #1258 ( MUL_MAT): ffn_up-34 ( 60K) [CUDA0 ]: blk.34.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-34 ( 15K) [CUDA0 ]
node #1259 ( MUL): ffn_gate_par-34 ( 60K) [CUDA0 ]: ffn_gelu-34 ( 60K) [CUDA0 ] ffn_up-34 ( 60K) [CUDA0 ]
node #1260 ( MUL_MAT): ffn_out-34 ( 15K) [CUDA0 ]: blk.34.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-34 ( 60K) [CUDA0 ]
node #1261 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-34 ( 15K) [CUDA0 ]
node #1262 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.34.post_ffw_norm ( 15K) [CUDA0 ]
node #1263 ( ADD): l_out-34 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-34 ( 15K) [CUDA0 ]
node #1264 ( RMS_NORM): norm-35 ( 15K) [CUDA0 ]: l_out-34 ( 15K) [CUDA0 ]
node #1265 ( MUL): attn_norm-35 ( 15K) [CUDA0 ]: norm-35 ( 15K) [CUDA0 ] blk.35.attn_norm.wei ( 15K) [CUDA0 ]
node #1266 ( MUL_MAT): Qcur-35 ( 16K) [CUDA0 ]: blk.35.attn_q.weight ( 8M) [CUDA0 ] attn_norm-35 ( 15K) [CUDA0 ]
node #1268 ( RMS_NORM): norm-35 ( 16K) [CUDA0 ]: Qcur-35 (reshaped) ( 16K) [CUDA0 ]
node #1269 ( MUL): Qcur_normed-35 ( 16K) [CUDA0 ]: norm-35 ( 16K) [CUDA0 ] blk.35.attn_q_norm.w ( 1K) [CUDA0 ]
node #1270 ( ROPE): Qcur-35 ( 16K) [CUDA0 ]: Qcur_normed-35 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1271 ( MUL_MAT): Kcur-35 ( 8K) [CUDA0 ]: blk.35.attn_k.weight ( 4M) [CUDA0 ] attn_norm-35 ( 15K) [CUDA0 ]
node #1273 ( RMS_NORM): norm-35 ( 8K) [CUDA0 ]: Kcur-35 (reshaped) ( 8K) [CUDA0 ]
node #1274 ( MUL): Kcur_normed-35 ( 8K) [CUDA0 ]: norm-35 ( 8K) [CUDA0 ] blk.35.attn_k_norm.w ( 1K) [CUDA0 ]
node #1275 ( ROPE): Kcur-35 ( 8K) [CUDA0 ]: Kcur_normed-35 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1276 ( MUL_MAT): Vcur-35 ( 8K) [CUDA0 ]: blk.35.attn_v.weight ( 6M) [CUDA0 ] attn_norm-35 ( 15K) [CUDA0 ]
node #1278 ( CPY): k_cache_view-35 (cop ( 2K) [CUDA0 ]: Kcur-35 ( 8K) [CUDA0 ] k_cache_view-35 ( 2K) [CUDA0 ]
node #1280 ( CPY): v_cache_view-35 (cop ( 2K) [CUDA0 ]: Vcur-35 ( 8K) [CUDA0 ] v_cache_view-35 ( 2K) [CUDA0 ]

SPLIT #72: CPU # 3 inputs: [q-35 ( 16K)] [k-35 ( 544K)] [v-35 ( 544K)]

node #1284 (FLASH_ATTN): node_1284 ( 16K) [ CPU ]: CPU#q-35#0 ( 16K) [ NULL ] CPU#k-35#0 ( 544K) [ NULL ] CPU#v-35#0 ( 544K) [ NULL ] CPU#KQ_mask (copy)#0 ( 32K) [ NULL ]

SPLIT #73: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1286 ( MUL_MAT): kqv_out-35 ( 15K) [CUDA0 ]: blk.35.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1287 ( RMS_NORM): norm-35 ( 15K) [CUDA0 ]: kqv_out-35 ( 15K) [CUDA0 ]
node #1288 ( MUL): attn_post_norm-35 ( 15K) [CUDA0 ]: norm-35 ( 15K) [CUDA0 ] blk.35.post_attentio ( 15K) [CUDA0 ]
node #1289 ( ADD): sa_out-35 ( 15K) [CUDA0 ]: attn_post_norm-35 ( 15K) [CUDA0 ] l_out-34 ( 15K) [CUDA0 ]
node #1290 ( RMS_NORM): norm-35 ( 15K) [CUDA0 ]: sa_out-35 ( 15K) [CUDA0 ]
node #1291 ( MUL): ffn_norm-35 ( 15K) [CUDA0 ]: norm-35 ( 15K) [CUDA0 ] blk.35.ffn_norm.weig ( 15K) [CUDA0 ]
node #1292 ( MUL_MAT): ffn_gate-35 ( 60K) [CUDA0 ]: blk.35.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-35 ( 15K) [CUDA0 ]
node #1293 ( UNARY): ffn_gelu-35 ( 60K) [CUDA0 ]: ffn_gate-35 ( 60K) [CUDA0 ]
node #1294 ( MUL_MAT): ffn_up-35 ( 60K) [CUDA0 ]: blk.35.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-35 ( 15K) [CUDA0 ]
node #1295 ( MUL): ffn_gate_par-35 ( 60K) [CUDA0 ]: ffn_gelu-35 ( 60K) [CUDA0 ] ffn_up-35 ( 60K) [CUDA0 ]
node #1296 ( MUL_MAT): ffn_out-35 ( 15K) [CUDA0 ]: blk.35.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-35 ( 60K) [CUDA0 ]
node #1297 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-35 ( 15K) [CUDA0 ]
node #1298 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.35.post_ffw_norm ( 15K) [CUDA0 ]
node #1299 ( ADD): l_out-35 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-35 ( 15K) [CUDA0 ]
node #1300 ( RMS_NORM): norm-36 ( 15K) [CUDA0 ]: l_out-35 ( 15K) [CUDA0 ]
node #1301 ( MUL): attn_norm-36 ( 15K) [CUDA0 ]: norm-36 ( 15K) [CUDA0 ] blk.36.attn_norm.wei ( 15K) [CUDA0 ]
node #1302 ( MUL_MAT): Qcur-36 ( 16K) [CUDA0 ]: blk.36.attn_q.weight ( 8M) [CUDA0 ] attn_norm-36 ( 15K) [CUDA0 ]
node #1304 ( RMS_NORM): norm-36 ( 16K) [CUDA0 ]: Qcur-36 (reshaped) ( 16K) [CUDA0 ]
node #1305 ( MUL): Qcur_normed-36 ( 16K) [CUDA0 ]: norm-36 ( 16K) [CUDA0 ] blk.36.attn_q_norm.w ( 1K) [CUDA0 ]
node #1306 ( ROPE): Qcur-36 ( 16K) [CUDA0 ]: Qcur_normed-36 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1307 ( MUL_MAT): Kcur-36 ( 8K) [CUDA0 ]: blk.36.attn_k.weight ( 4M) [CUDA0 ] attn_norm-36 ( 15K) [CUDA0 ]
node #1309 ( RMS_NORM): norm-36 ( 8K) [CUDA0 ]: Kcur-36 (reshaped) ( 8K) [CUDA0 ]
node #1310 ( MUL): Kcur_normed-36 ( 8K) [CUDA0 ]: norm-36 ( 8K) [CUDA0 ] blk.36.attn_k_norm.w ( 1K) [CUDA0 ]
node #1311 ( ROPE): Kcur-36 ( 8K) [CUDA0 ]: Kcur_normed-36 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1312 ( MUL_MAT): Vcur-36 ( 8K) [CUDA0 ]: blk.36.attn_v.weight ( 4M) [CUDA0 ] attn_norm-36 ( 15K) [CUDA0 ]
node #1314 ( CPY): k_cache_view-36 (cop ( 2K) [CUDA0 ]: Kcur-36 ( 8K) [CUDA0 ] k_cache_view-36 ( 2K) [CUDA0 ]
node #1316 ( CPY): v_cache_view-36 (cop ( 2K) [CUDA0 ]: Vcur-36 ( 8K) [CUDA0 ] v_cache_view-36 ( 2K) [CUDA0 ]

SPLIT #74: CPU # 3 inputs: [q-36 ( 16K)] [k-36 ( 544K)] [v-36 ( 544K)]

node #1320 (FLASH_ATTN): node_1320 ( 16K) [ CPU ]: CPU#q-36#0 ( 16K) [ NULL ] CPU#k-36#0 ( 544K) [ NULL ] CPU#v-36#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #75: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1322 ( MUL_MAT): kqv_out-36 ( 15K) [CUDA0 ]: blk.36.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1323 ( RMS_NORM): norm-36 ( 15K) [CUDA0 ]: kqv_out-36 ( 15K) [CUDA0 ]
node #1324 ( MUL): attn_post_norm-36 ( 15K) [CUDA0 ]: norm-36 ( 15K) [CUDA0 ] blk.36.post_attentio ( 15K) [CUDA0 ]
node #1325 ( ADD): sa_out-36 ( 15K) [CUDA0 ]: attn_post_norm-36 ( 15K) [CUDA0 ] l_out-35 ( 15K) [CUDA0 ]
node #1326 ( RMS_NORM): norm-36 ( 15K) [CUDA0 ]: sa_out-36 ( 15K) [CUDA0 ]
node #1327 ( MUL): ffn_norm-36 ( 15K) [CUDA0 ]: norm-36 ( 15K) [CUDA0 ] blk.36.ffn_norm.weig ( 15K) [CUDA0 ]
node #1328 ( MUL_MAT): ffn_gate-36 ( 60K) [CUDA0 ]: blk.36.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-36 ( 15K) [CUDA0 ]
node #1329 ( UNARY): ffn_gelu-36 ( 60K) [CUDA0 ]: ffn_gate-36 ( 60K) [CUDA0 ]
node #1330 ( MUL_MAT): ffn_up-36 ( 60K) [CUDA0 ]: blk.36.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-36 ( 15K) [CUDA0 ]
node #1331 ( MUL): ffn_gate_par-36 ( 60K) [CUDA0 ]: ffn_gelu-36 ( 60K) [CUDA0 ] ffn_up-36 ( 60K) [CUDA0 ]
node #1332 ( MUL_MAT): ffn_out-36 ( 15K) [CUDA0 ]: blk.36.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-36 ( 60K) [CUDA0 ]
node #1333 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-36 ( 15K) [CUDA0 ]
node #1334 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.36.post_ffw_norm ( 15K) [CUDA0 ]
node #1335 ( ADD): l_out-36 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-36 ( 15K) [CUDA0 ]
node #1336 ( RMS_NORM): norm-37 ( 15K) [CUDA0 ]: l_out-36 ( 15K) [CUDA0 ]
node #1337 ( MUL): attn_norm-37 ( 15K) [CUDA0 ]: norm-37 ( 15K) [CUDA0 ] blk.37.attn_norm.wei ( 15K) [CUDA0 ]
node #1338 ( MUL_MAT): Qcur-37 ( 16K) [CUDA0 ]: blk.37.attn_q.weight ( 8M) [CUDA0 ] attn_norm-37 ( 15K) [CUDA0 ]
node #1340 ( RMS_NORM): norm-37 ( 16K) [CUDA0 ]: Qcur-37 (reshaped) ( 16K) [CUDA0 ]
node #1341 ( MUL): Qcur_normed-37 ( 16K) [CUDA0 ]: norm-37 ( 16K) [CUDA0 ] blk.37.attn_q_norm.w ( 1K) [CUDA0 ]
node #1342 ( ROPE): Qcur-37 ( 16K) [CUDA0 ]: Qcur_normed-37 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1343 ( MUL_MAT): Kcur-37 ( 8K) [CUDA0 ]: blk.37.attn_k.weight ( 4M) [CUDA0 ] attn_norm-37 ( 15K) [CUDA0 ]
node #1345 ( RMS_NORM): norm-37 ( 8K) [CUDA0 ]: Kcur-37 (reshaped) ( 8K) [CUDA0 ]
node #1346 ( MUL): Kcur_normed-37 ( 8K) [CUDA0 ]: norm-37 ( 8K) [CUDA0 ] blk.37.attn_k_norm.w ( 1K) [CUDA0 ]
node #1347 ( ROPE): Kcur-37 ( 8K) [CUDA0 ]: Kcur_normed-37 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1348 ( MUL_MAT): Vcur-37 ( 8K) [CUDA0 ]: blk.37.attn_v.weight ( 4M) [CUDA0 ] attn_norm-37 ( 15K) [CUDA0 ]
node #1350 ( CPY): k_cache_view-37 (cop ( 2K) [CUDA0 ]: Kcur-37 ( 8K) [CUDA0 ] k_cache_view-37 ( 2K) [CUDA0 ]
node #1352 ( CPY): v_cache_view-37 (cop ( 2K) [CUDA0 ]: Vcur-37 ( 8K) [CUDA0 ] v_cache_view-37 ( 2K) [CUDA0 ]

SPLIT #76: CPU # 3 inputs: [q-37 ( 16K)] [k-37 ( 544K)] [v-37 ( 544K)]

node #1356 (FLASH_ATTN): node_1356 ( 16K) [ CPU ]: CPU#q-37#0 ( 16K) [ NULL ] CPU#k-37#0 ( 544K) [ NULL ] CPU#v-37#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #77: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1358 ( MUL_MAT): kqv_out-37 ( 15K) [CUDA0 ]: blk.37.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1359 ( RMS_NORM): norm-37 ( 15K) [CUDA0 ]: kqv_out-37 ( 15K) [CUDA0 ]
node #1360 ( MUL): attn_post_norm-37 ( 15K) [CUDA0 ]: norm-37 ( 15K) [CUDA0 ] blk.37.post_attentio ( 15K) [CUDA0 ]
node #1361 ( ADD): sa_out-37 ( 15K) [CUDA0 ]: attn_post_norm-37 ( 15K) [CUDA0 ] l_out-36 ( 15K) [CUDA0 ]
node #1362 ( RMS_NORM): norm-37 ( 15K) [CUDA0 ]: sa_out-37 ( 15K) [CUDA0 ]
node #1363 ( MUL): ffn_norm-37 ( 15K) [CUDA0 ]: norm-37 ( 15K) [CUDA0 ] blk.37.ffn_norm.weig ( 15K) [CUDA0 ]
node #1364 ( MUL_MAT): ffn_gate-37 ( 60K) [CUDA0 ]: blk.37.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-37 ( 15K) [CUDA0 ]
node #1365 ( UNARY): ffn_gelu-37 ( 60K) [CUDA0 ]: ffn_gate-37 ( 60K) [CUDA0 ]
node #1366 ( MUL_MAT): ffn_up-37 ( 60K) [CUDA0 ]: blk.37.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-37 ( 15K) [CUDA0 ]
node #1367 ( MUL): ffn_gate_par-37 ( 60K) [CUDA0 ]: ffn_gelu-37 ( 60K) [CUDA0 ] ffn_up-37 ( 60K) [CUDA0 ]
node #1368 ( MUL_MAT): ffn_out-37 ( 15K) [CUDA0 ]: blk.37.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-37 ( 60K) [CUDA0 ]
node #1369 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-37 ( 15K) [CUDA0 ]
node #1370 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.37.post_ffw_norm ( 15K) [CUDA0 ]
node #1371 ( ADD): l_out-37 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-37 ( 15K) [CUDA0 ]
node #1372 ( RMS_NORM): norm-38 ( 15K) [CUDA0 ]: l_out-37 ( 15K) [CUDA0 ]
node #1373 ( MUL): attn_norm-38 ( 15K) [CUDA0 ]: norm-38 ( 15K) [CUDA0 ] blk.38.attn_norm.wei ( 15K) [CUDA0 ]
node #1374 ( MUL_MAT): Qcur-38 ( 16K) [CUDA0 ]: blk.38.attn_q.weight ( 8M) [CUDA0 ] attn_norm-38 ( 15K) [CUDA0 ]
node #1376 ( RMS_NORM): norm-38 ( 16K) [CUDA0 ]: Qcur-38 (reshaped) ( 16K) [CUDA0 ]
node #1377 ( MUL): Qcur_normed-38 ( 16K) [CUDA0 ]: norm-38 ( 16K) [CUDA0 ] blk.38.attn_q_norm.w ( 1K) [CUDA0 ]
node #1378 ( ROPE): Qcur-38 ( 16K) [CUDA0 ]: Qcur_normed-38 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1379 ( MUL_MAT): Kcur-38 ( 8K) [CUDA0 ]: blk.38.attn_k.weight ( 4M) [CUDA0 ] attn_norm-38 ( 15K) [CUDA0 ]
node #1381 ( RMS_NORM): norm-38 ( 8K) [CUDA0 ]: Kcur-38 (reshaped) ( 8K) [CUDA0 ]
node #1382 ( MUL): Kcur_normed-38 ( 8K) [CUDA0 ]: norm-38 ( 8K) [CUDA0 ] blk.38.attn_k_norm.w ( 1K) [CUDA0 ]
node #1383 ( ROPE): Kcur-38 ( 8K) [CUDA0 ]: Kcur_normed-38 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1384 ( MUL_MAT): Vcur-38 ( 8K) [CUDA0 ]: blk.38.attn_v.weight ( 6M) [CUDA0 ] attn_norm-38 ( 15K) [CUDA0 ]
node #1386 ( CPY): k_cache_view-38 (cop ( 2K) [CUDA0 ]: Kcur-38 ( 8K) [CUDA0 ] k_cache_view-38 ( 2K) [CUDA0 ]
node #1388 ( CPY): v_cache_view-38 (cop ( 2K) [CUDA0 ]: Vcur-38 ( 8K) [CUDA0 ] v_cache_view-38 ( 2K) [CUDA0 ]

SPLIT #78: CPU # 3 inputs: [q-38 ( 16K)] [k-38 ( 544K)] [v-38 ( 544K)]

node #1392 (FLASH_ATTN): node_1392 ( 16K) [ CPU ]: CPU#q-38#0 ( 16K) [ NULL ] CPU#k-38#0 ( 544K) [ NULL ] CPU#v-38#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #79: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1394 ( MUL_MAT): kqv_out-38 ( 15K) [CUDA0 ]: blk.38.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1395 ( RMS_NORM): norm-38 ( 15K) [CUDA0 ]: kqv_out-38 ( 15K) [CUDA0 ]
node #1396 ( MUL): attn_post_norm-38 ( 15K) [CUDA0 ]: norm-38 ( 15K) [CUDA0 ] blk.38.post_attentio ( 15K) [CUDA0 ]
node #1397 ( ADD): sa_out-38 ( 15K) [CUDA0 ]: attn_post_norm-38 ( 15K) [CUDA0 ] l_out-37 ( 15K) [CUDA0 ]
node #1398 ( RMS_NORM): norm-38 ( 15K) [CUDA0 ]: sa_out-38 ( 15K) [CUDA0 ]
node #1399 ( MUL): ffn_norm-38 ( 15K) [CUDA0 ]: norm-38 ( 15K) [CUDA0 ] blk.38.ffn_norm.weig ( 15K) [CUDA0 ]
node #1400 ( MUL_MAT): ffn_gate-38 ( 60K) [CUDA0 ]: blk.38.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-38 ( 15K) [CUDA0 ]
node #1401 ( UNARY): ffn_gelu-38 ( 60K) [CUDA0 ]: ffn_gate-38 ( 60K) [CUDA0 ]
node #1402 ( MUL_MAT): ffn_up-38 ( 60K) [CUDA0 ]: blk.38.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-38 ( 15K) [CUDA0 ]
node #1403 ( MUL): ffn_gate_par-38 ( 60K) [CUDA0 ]: ffn_gelu-38 ( 60K) [CUDA0 ] ffn_up-38 ( 60K) [CUDA0 ]
node #1404 ( MUL_MAT): ffn_out-38 ( 15K) [CUDA0 ]: blk.38.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-38 ( 60K) [CUDA0 ]
node #1405 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-38 ( 15K) [CUDA0 ]
node #1406 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.38.post_ffw_norm ( 15K) [CUDA0 ]
node #1407 ( ADD): l_out-38 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-38 ( 15K) [CUDA0 ]
node #1408 ( RMS_NORM): norm-39 ( 15K) [CUDA0 ]: l_out-38 ( 15K) [CUDA0 ]
node #1409 ( MUL): attn_norm-39 ( 15K) [CUDA0 ]: norm-39 ( 15K) [CUDA0 ] blk.39.attn_norm.wei ( 15K) [CUDA0 ]
node #1410 ( MUL_MAT): Qcur-39 ( 16K) [CUDA0 ]: blk.39.attn_q.weight ( 8M) [CUDA0 ] attn_norm-39 ( 15K) [CUDA0 ]
node #1412 ( RMS_NORM): norm-39 ( 16K) [CUDA0 ]: Qcur-39 (reshaped) ( 16K) [CUDA0 ]
node #1413 ( MUL): Qcur_normed-39 ( 16K) [CUDA0 ]: norm-39 ( 16K) [CUDA0 ] blk.39.attn_q_norm.w ( 1K) [CUDA0 ]
node #1414 ( ROPE): Qcur-39 ( 16K) [CUDA0 ]: Qcur_normed-39 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1415 ( MUL_MAT): Kcur-39 ( 8K) [CUDA0 ]: blk.39.attn_k.weight ( 4M) [CUDA0 ] attn_norm-39 ( 15K) [CUDA0 ]
node #1417 ( RMS_NORM): norm-39 ( 8K) [CUDA0 ]: Kcur-39 (reshaped) ( 8K) [CUDA0 ]
node #1418 ( MUL): Kcur_normed-39 ( 8K) [CUDA0 ]: norm-39 ( 8K) [CUDA0 ] blk.39.attn_k_norm.w ( 1K) [CUDA0 ]
node #1419 ( ROPE): Kcur-39 ( 8K) [CUDA0 ]: Kcur_normed-39 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1420 ( MUL_MAT): Vcur-39 ( 8K) [CUDA0 ]: blk.39.attn_v.weight ( 4M) [CUDA0 ] attn_norm-39 ( 15K) [CUDA0 ]
node #1422 ( CPY): k_cache_view-39 (cop ( 2K) [CUDA0 ]: Kcur-39 ( 8K) [CUDA0 ] k_cache_view-39 ( 2K) [CUDA0 ]
node #1424 ( CPY): v_cache_view-39 (cop ( 2K) [CUDA0 ]: Vcur-39 ( 8K) [CUDA0 ] v_cache_view-39 ( 2K) [CUDA0 ]

SPLIT #80: CPU # 3 inputs: [q-39 ( 16K)] [k-39 ( 544K)] [v-39 ( 544K)]

node #1428 (FLASH_ATTN): node_1428 ( 16K) [ CPU ]: CPU#q-39#0 ( 16K) [ NULL ] CPU#k-39#0 ( 544K) [ NULL ] CPU#v-39#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #81: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1430 ( MUL_MAT): kqv_out-39 ( 15K) [CUDA0 ]: blk.39.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1431 ( RMS_NORM): norm-39 ( 15K) [CUDA0 ]: kqv_out-39 ( 15K) [CUDA0 ]
node #1432 ( MUL): attn_post_norm-39 ( 15K) [CUDA0 ]: norm-39 ( 15K) [CUDA0 ] blk.39.post_attentio ( 15K) [CUDA0 ]
node #1433 ( ADD): sa_out-39 ( 15K) [CUDA0 ]: attn_post_norm-39 ( 15K) [CUDA0 ] l_out-38 ( 15K) [CUDA0 ]
node #1434 ( RMS_NORM): norm-39 ( 15K) [CUDA0 ]: sa_out-39 ( 15K) [CUDA0 ]
node #1435 ( MUL): ffn_norm-39 ( 15K) [CUDA0 ]: norm-39 ( 15K) [CUDA0 ] blk.39.ffn_norm.weig ( 15K) [CUDA0 ]
node #1436 ( MUL_MAT): ffn_gate-39 ( 60K) [CUDA0 ]: blk.39.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-39 ( 15K) [CUDA0 ]
node #1437 ( UNARY): ffn_gelu-39 ( 60K) [CUDA0 ]: ffn_gate-39 ( 60K) [CUDA0 ]
node #1438 ( MUL_MAT): ffn_up-39 ( 60K) [CUDA0 ]: blk.39.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-39 ( 15K) [CUDA0 ]
node #1439 ( MUL): ffn_gate_par-39 ( 60K) [CUDA0 ]: ffn_gelu-39 ( 60K) [CUDA0 ] ffn_up-39 ( 60K) [CUDA0 ]
node #1440 ( MUL_MAT): ffn_out-39 ( 15K) [CUDA0 ]: blk.39.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-39 ( 60K) [CUDA0 ]
node #1441 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-39 ( 15K) [CUDA0 ]
node #1442 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.39.post_ffw_norm ( 15K) [CUDA0 ]
node #1443 ( ADD): l_out-39 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-39 ( 15K) [CUDA0 ]
node #1444 ( RMS_NORM): norm-40 ( 15K) [CUDA0 ]: l_out-39 ( 15K) [CUDA0 ]
node #1445 ( MUL): attn_norm-40 ( 15K) [CUDA0 ]: norm-40 ( 15K) [CUDA0 ] blk.40.attn_norm.wei ( 15K) [CUDA0 ]
node #1446 ( MUL_MAT): Qcur-40 ( 16K) [CUDA0 ]: blk.40.attn_q.weight ( 8M) [CUDA0 ] attn_norm-40 ( 15K) [CUDA0 ]
node #1448 ( RMS_NORM): norm-40 ( 16K) [CUDA0 ]: Qcur-40 (reshaped) ( 16K) [CUDA0 ]
node #1449 ( MUL): Qcur_normed-40 ( 16K) [CUDA0 ]: norm-40 ( 16K) [CUDA0 ] blk.40.attn_q_norm.w ( 1K) [CUDA0 ]
node #1450 ( ROPE): Qcur-40 ( 16K) [CUDA0 ]: Qcur_normed-40 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1451 ( MUL_MAT): Kcur-40 ( 8K) [CUDA0 ]: blk.40.attn_k.weight ( 4M) [CUDA0 ] attn_norm-40 ( 15K) [CUDA0 ]
node #1453 ( RMS_NORM): norm-40 ( 8K) [CUDA0 ]: Kcur-40 (reshaped) ( 8K) [CUDA0 ]
node #1454 ( MUL): Kcur_normed-40 ( 8K) [CUDA0 ]: norm-40 ( 8K) [CUDA0 ] blk.40.attn_k_norm.w ( 1K) [CUDA0 ]
node #1455 ( ROPE): Kcur-40 ( 8K) [CUDA0 ]: Kcur_normed-40 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1456 ( MUL_MAT): Vcur-40 ( 8K) [CUDA0 ]: blk.40.attn_v.weight ( 4M) [CUDA0 ] attn_norm-40 ( 15K) [CUDA0 ]
node #1458 ( CPY): k_cache_view-40 (cop ( 2K) [CUDA0 ]: Kcur-40 ( 8K) [CUDA0 ] k_cache_view-40 ( 2K) [CUDA0 ]
node #1460 ( CPY): v_cache_view-40 (cop ( 2K) [CUDA0 ]: Vcur-40 ( 8K) [CUDA0 ] v_cache_view-40 ( 2K) [CUDA0 ]

SPLIT #82: CPU # 3 inputs: [q-40 ( 16K)] [k-40 ( 544K)] [v-40 ( 544K)]

node #1464 (FLASH_ATTN): node_1464 ( 16K) [ CPU ]: CPU#q-40#0 ( 16K) [ NULL ] CPU#k-40#0 ( 544K) [ NULL ] CPU#v-40#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #83: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1466 ( MUL_MAT): kqv_out-40 ( 15K) [CUDA0 ]: blk.40.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1467 ( RMS_NORM): norm-40 ( 15K) [CUDA0 ]: kqv_out-40 ( 15K) [CUDA0 ]
node #1468 ( MUL): attn_post_norm-40 ( 15K) [CUDA0 ]: norm-40 ( 15K) [CUDA0 ] blk.40.post_attentio ( 15K) [CUDA0 ]
node #1469 ( ADD): sa_out-40 ( 15K) [CUDA0 ]: attn_post_norm-40 ( 15K) [CUDA0 ] l_out-39 ( 15K) [CUDA0 ]
node #1470 ( RMS_NORM): norm-40 ( 15K) [CUDA0 ]: sa_out-40 ( 15K) [CUDA0 ]
node #1471 ( MUL): ffn_norm-40 ( 15K) [CUDA0 ]: norm-40 ( 15K) [CUDA0 ] blk.40.ffn_norm.weig ( 15K) [CUDA0 ]
node #1472 ( MUL_MAT): ffn_gate-40 ( 60K) [CUDA0 ]: blk.40.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-40 ( 15K) [CUDA0 ]
node #1473 ( UNARY): ffn_gelu-40 ( 60K) [CUDA0 ]: ffn_gate-40 ( 60K) [CUDA0 ]
node #1474 ( MUL_MAT): ffn_up-40 ( 60K) [CUDA0 ]: blk.40.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-40 ( 15K) [CUDA0 ]
node #1475 ( MUL): ffn_gate_par-40 ( 60K) [CUDA0 ]: ffn_gelu-40 ( 60K) [CUDA0 ] ffn_up-40 ( 60K) [CUDA0 ]
node #1476 ( MUL_MAT): ffn_out-40 ( 15K) [CUDA0 ]: blk.40.ffn_down.weig ( 31M) [CUDA0 ] ffn_gate_par-40 ( 60K) [CUDA0 ]
node #1477 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-40 ( 15K) [CUDA0 ]
node #1478 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.40.post_ffw_norm ( 15K) [CUDA0 ]
node #1479 ( ADD): l_out-40 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-40 ( 15K) [CUDA0 ]
node #1480 ( RMS_NORM): norm-41 ( 15K) [CUDA0 ]: l_out-40 ( 15K) [CUDA0 ]
node #1481 ( MUL): attn_norm-41 ( 15K) [CUDA0 ]: norm-41 ( 15K) [CUDA0 ] blk.41.attn_norm.wei ( 15K) [CUDA0 ]
node #1482 ( MUL_MAT): Qcur-41 ( 16K) [CUDA0 ]: blk.41.attn_q.weight ( 8M) [CUDA0 ] attn_norm-41 ( 15K) [CUDA0 ]
node #1484 ( RMS_NORM): norm-41 ( 16K) [CUDA0 ]: Qcur-41 (reshaped) ( 16K) [CUDA0 ]
node #1485 ( MUL): Qcur_normed-41 ( 16K) [CUDA0 ]: norm-41 ( 16K) [CUDA0 ] blk.41.attn_q_norm.w ( 1K) [CUDA0 ]
node #1486 ( ROPE): Qcur-41 ( 16K) [CUDA0 ]: Qcur_normed-41 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1487 ( MUL_MAT): Kcur-41 ( 8K) [CUDA0 ]: blk.41.attn_k.weight ( 4M) [CUDA0 ] attn_norm-41 ( 15K) [CUDA0 ]
node #1489 ( RMS_NORM): norm-41 ( 8K) [CUDA0 ]: Kcur-41 (reshaped) ( 8K) [CUDA0 ]
node #1490 ( MUL): Kcur_normed-41 ( 8K) [CUDA0 ]: norm-41 ( 8K) [CUDA0 ] blk.41.attn_k_norm.w ( 1K) [CUDA0 ]
node #1491 ( ROPE): Kcur-41 ( 8K) [CUDA0 ]: Kcur_normed-41 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1492 ( MUL_MAT): Vcur-41 ( 8K) [CUDA0 ]: blk.41.attn_v.weight ( 6M) [CUDA0 ] attn_norm-41 ( 15K) [CUDA0 ]
node #1494 ( CPY): k_cache_view-41 (cop ( 2K) [CUDA0 ]: Kcur-41 ( 8K) [CUDA0 ] k_cache_view-41 ( 2K) [CUDA0 ]
node #1496 ( CPY): v_cache_view-41 (cop ( 2K) [CUDA0 ]: Vcur-41 ( 8K) [CUDA0 ] v_cache_view-41 ( 2K) [CUDA0 ]

SPLIT #84: CPU # 3 inputs: [q-41 ( 16K)] [k-41 ( 544K)] [v-41 ( 544K)]

node #1500 (FLASH_ATTN): node_1500 ( 16K) [ CPU ]: CPU#q-41#0 ( 16K) [ NULL ] CPU#k-41#0 ( 544K) [ NULL ] CPU#v-41#0 ( 544K) [ NULL ] CPU#KQ_mask (copy)#0 ( 32K) [ NULL ]

SPLIT #85: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1502 ( MUL_MAT): kqv_out-41 ( 15K) [CUDA0 ]: blk.41.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1503 ( RMS_NORM): norm-41 ( 15K) [CUDA0 ]: kqv_out-41 ( 15K) [CUDA0 ]
node #1504 ( MUL): attn_post_norm-41 ( 15K) [CUDA0 ]: norm-41 ( 15K) [CUDA0 ] blk.41.post_attentio ( 15K) [CUDA0 ]
node #1505 ( ADD): sa_out-41 ( 15K) [CUDA0 ]: attn_post_norm-41 ( 15K) [CUDA0 ] l_out-40 ( 15K) [CUDA0 ]
node #1506 ( RMS_NORM): norm-41 ( 15K) [CUDA0 ]: sa_out-41 ( 15K) [CUDA0 ]
node #1507 ( MUL): ffn_norm-41 ( 15K) [CUDA0 ]: norm-41 ( 15K) [CUDA0 ] blk.41.ffn_norm.weig ( 15K) [CUDA0 ]
node #1508 ( MUL_MAT): ffn_gate-41 ( 60K) [CUDA0 ]: blk.41.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-41 ( 15K) [CUDA0 ]
node #1509 ( UNARY): ffn_gelu-41 ( 60K) [CUDA0 ]: ffn_gate-41 ( 60K) [CUDA0 ]
node #1510 ( MUL_MAT): ffn_up-41 ( 60K) [CUDA0 ]: blk.41.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-41 ( 15K) [CUDA0 ]
node #1511 ( MUL): ffn_gate_par-41 ( 60K) [CUDA0 ]: ffn_gelu-41 ( 60K) [CUDA0 ] ffn_up-41 ( 60K) [CUDA0 ]
node #1512 ( MUL_MAT): ffn_out-41 ( 15K) [CUDA0 ]: blk.41.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-41 ( 60K) [CUDA0 ]
node #1513 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-41 ( 15K) [CUDA0 ]
node #1514 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.41.post_ffw_norm ( 15K) [CUDA0 ]
node #1515 ( ADD): l_out-41 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-41 ( 15K) [CUDA0 ]
node #1516 ( RMS_NORM): norm-42 ( 15K) [CUDA0 ]: l_out-41 ( 15K) [CUDA0 ]
node #1517 ( MUL): attn_norm-42 ( 15K) [CUDA0 ]: norm-42 ( 15K) [CUDA0 ] blk.42.attn_norm.wei ( 15K) [CUDA0 ]
node #1518 ( MUL_MAT): Qcur-42 ( 16K) [CUDA0 ]: blk.42.attn_q.weight ( 8M) [CUDA0 ] attn_norm-42 ( 15K) [CUDA0 ]
node #1520 ( RMS_NORM): norm-42 ( 16K) [CUDA0 ]: Qcur-42 (reshaped) ( 16K) [CUDA0 ]
node #1521 ( MUL): Qcur_normed-42 ( 16K) [CUDA0 ]: norm-42 ( 16K) [CUDA0 ] blk.42.attn_q_norm.w ( 1K) [CUDA0 ]
node #1522 ( ROPE): Qcur-42 ( 16K) [CUDA0 ]: Qcur_normed-42 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1523 ( MUL_MAT): Kcur-42 ( 8K) [CUDA0 ]: blk.42.attn_k.weight ( 4M) [CUDA0 ] attn_norm-42 ( 15K) [CUDA0 ]
node #1525 ( RMS_NORM): norm-42 ( 8K) [CUDA0 ]: Kcur-42 (reshaped) ( 8K) [CUDA0 ]
node #1526 ( MUL): Kcur_normed-42 ( 8K) [CUDA0 ]: norm-42 ( 8K) [CUDA0 ] blk.42.attn_k_norm.w ( 1K) [CUDA0 ]
node #1527 ( ROPE): Kcur-42 ( 8K) [CUDA0 ]: Kcur_normed-42 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1528 ( MUL_MAT): Vcur-42 ( 8K) [CUDA0 ]: blk.42.attn_v.weight ( 6M) [CUDA0 ] attn_norm-42 ( 15K) [CUDA0 ]
node #1530 ( CPY): k_cache_view-42 (cop ( 2K) [CUDA0 ]: Kcur-42 ( 8K) [CUDA0 ] k_cache_view-42 ( 2K) [CUDA0 ]
node #1532 ( CPY): v_cache_view-42 (cop ( 2K) [CUDA0 ]: Vcur-42 ( 8K) [CUDA0 ] v_cache_view-42 ( 2K) [CUDA0 ]

SPLIT #86: CPU # 3 inputs: [q-42 ( 16K)] [k-42 ( 544K)] [v-42 ( 544K)]

node #1536 (FLASH_ATTN): node_1536 ( 16K) [ CPU ]: CPU#q-42#0 ( 16K) [ NULL ] CPU#k-42#0 ( 544K) [ NULL ] CPU#v-42#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #87: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1538 ( MUL_MAT): kqv_out-42 ( 15K) [CUDA0 ]: blk.42.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1539 ( RMS_NORM): norm-42 ( 15K) [CUDA0 ]: kqv_out-42 ( 15K) [CUDA0 ]
node #1540 ( MUL): attn_post_norm-42 ( 15K) [CUDA0 ]: norm-42 ( 15K) [CUDA0 ] blk.42.post_attentio ( 15K) [CUDA0 ]
node #1541 ( ADD): sa_out-42 ( 15K) [CUDA0 ]: attn_post_norm-42 ( 15K) [CUDA0 ] l_out-41 ( 15K) [CUDA0 ]
node #1542 ( RMS_NORM): norm-42 ( 15K) [CUDA0 ]: sa_out-42 ( 15K) [CUDA0 ]
node #1543 ( MUL): ffn_norm-42 ( 15K) [CUDA0 ]: norm-42 ( 15K) [CUDA0 ] blk.42.ffn_norm.weig ( 15K) [CUDA0 ]
node #1544 ( MUL_MAT): ffn_gate-42 ( 60K) [CUDA0 ]: blk.42.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-42 ( 15K) [CUDA0 ]
node #1545 ( UNARY): ffn_gelu-42 ( 60K) [CUDA0 ]: ffn_gate-42 ( 60K) [CUDA0 ]
node #1546 ( MUL_MAT): ffn_up-42 ( 60K) [CUDA0 ]: blk.42.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-42 ( 15K) [CUDA0 ]
node #1547 ( MUL): ffn_gate_par-42 ( 60K) [CUDA0 ]: ffn_gelu-42 ( 60K) [CUDA0 ] ffn_up-42 ( 60K) [CUDA0 ]
node #1548 ( MUL_MAT): ffn_out-42 ( 15K) [CUDA0 ]: blk.42.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-42 ( 60K) [CUDA0 ]
node #1549 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-42 ( 15K) [CUDA0 ]
node #1550 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.42.post_ffw_norm ( 15K) [CUDA0 ]
node #1551 ( ADD): l_out-42 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-42 ( 15K) [CUDA0 ]
node #1552 ( RMS_NORM): norm-43 ( 15K) [CUDA0 ]: l_out-42 ( 15K) [CUDA0 ]
node #1553 ( MUL): attn_norm-43 ( 15K) [CUDA0 ]: norm-43 ( 15K) [CUDA0 ] blk.43.attn_norm.wei ( 15K) [CUDA0 ]
node #1554 ( MUL_MAT): Qcur-43 ( 16K) [CUDA0 ]: blk.43.attn_q.weight ( 8M) [CUDA0 ] attn_norm-43 ( 15K) [CUDA0 ]
node #1556 ( RMS_NORM): norm-43 ( 16K) [CUDA0 ]: Qcur-43 (reshaped) ( 16K) [CUDA0 ]
node #1557 ( MUL): Qcur_normed-43 ( 16K) [CUDA0 ]: norm-43 ( 16K) [CUDA0 ] blk.43.attn_q_norm.w ( 1K) [CUDA0 ]
node #1558 ( ROPE): Qcur-43 ( 16K) [CUDA0 ]: Qcur_normed-43 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1559 ( MUL_MAT): Kcur-43 ( 8K) [CUDA0 ]: blk.43.attn_k.weight ( 4M) [CUDA0 ] attn_norm-43 ( 15K) [CUDA0 ]
node #1561 ( RMS_NORM): norm-43 ( 8K) [CUDA0 ]: Kcur-43 (reshaped) ( 8K) [CUDA0 ]
node #1562 ( MUL): Kcur_normed-43 ( 8K) [CUDA0 ]: norm-43 ( 8K) [CUDA0 ] blk.43.attn_k_norm.w ( 1K) [CUDA0 ]
node #1563 ( ROPE): Kcur-43 ( 8K) [CUDA0 ]: Kcur_normed-43 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1564 ( MUL_MAT): Vcur-43 ( 8K) [CUDA0 ]: blk.43.attn_v.weight ( 6M) [CUDA0 ] attn_norm-43 ( 15K) [CUDA0 ]
node #1566 ( CPY): k_cache_view-43 (cop ( 2K) [CUDA0 ]: Kcur-43 ( 8K) [CUDA0 ] k_cache_view-43 ( 2K) [CUDA0 ]
node #1568 ( CPY): v_cache_view-43 (cop ( 2K) [CUDA0 ]: Vcur-43 ( 8K) [CUDA0 ] v_cache_view-43 ( 2K) [CUDA0 ]

SPLIT #88: CPU # 3 inputs: [q-43 ( 16K)] [k-43 ( 544K)] [v-43 ( 544K)]

node #1572 (FLASH_ATTN): node_1572 ( 16K) [ CPU ]: CPU#q-43#0 ( 16K) [ NULL ] CPU#k-43#0 ( 544K) [ NULL ] CPU#v-43#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #89: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1574 ( MUL_MAT): kqv_out-43 ( 15K) [CUDA0 ]: blk.43.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1575 ( RMS_NORM): norm-43 ( 15K) [CUDA0 ]: kqv_out-43 ( 15K) [CUDA0 ]
node #1576 ( MUL): attn_post_norm-43 ( 15K) [CUDA0 ]: norm-43 ( 15K) [CUDA0 ] blk.43.post_attentio ( 15K) [CUDA0 ]
node #1577 ( ADD): sa_out-43 ( 15K) [CUDA0 ]: attn_post_norm-43 ( 15K) [CUDA0 ] l_out-42 ( 15K) [CUDA0 ]
node #1578 ( RMS_NORM): norm-43 ( 15K) [CUDA0 ]: sa_out-43 ( 15K) [CUDA0 ]
node #1579 ( MUL): ffn_norm-43 ( 15K) [CUDA0 ]: norm-43 ( 15K) [CUDA0 ] blk.43.ffn_norm.weig ( 15K) [CUDA0 ]
node #1580 ( MUL_MAT): ffn_gate-43 ( 60K) [CUDA0 ]: blk.43.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-43 ( 15K) [CUDA0 ]
node #1581 ( UNARY): ffn_gelu-43 ( 60K) [CUDA0 ]: ffn_gate-43 ( 60K) [CUDA0 ]
node #1582 ( MUL_MAT): ffn_up-43 ( 60K) [CUDA0 ]: blk.43.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-43 ( 15K) [CUDA0 ]
node #1583 ( MUL): ffn_gate_par-43 ( 60K) [CUDA0 ]: ffn_gelu-43 ( 60K) [CUDA0 ] ffn_up-43 ( 60K) [CUDA0 ]
node #1584 ( MUL_MAT): ffn_out-43 ( 15K) [CUDA0 ]: blk.43.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-43 ( 60K) [CUDA0 ]
node #1585 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-43 ( 15K) [CUDA0 ]
node #1586 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.43.post_ffw_norm ( 15K) [CUDA0 ]
node #1587 ( ADD): l_out-43 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-43 ( 15K) [CUDA0 ]
node #1588 ( RMS_NORM): norm-44 ( 15K) [CUDA0 ]: l_out-43 ( 15K) [CUDA0 ]
node #1589 ( MUL): attn_norm-44 ( 15K) [CUDA0 ]: norm-44 ( 15K) [CUDA0 ] blk.44.attn_norm.wei ( 15K) [CUDA0 ]
node #1590 ( MUL_MAT): Qcur-44 ( 16K) [CUDA0 ]: blk.44.attn_q.weight ( 8M) [CUDA0 ] attn_norm-44 ( 15K) [CUDA0 ]
node #1592 ( RMS_NORM): norm-44 ( 16K) [CUDA0 ]: Qcur-44 (reshaped) ( 16K) [CUDA0 ]
node #1593 ( MUL): Qcur_normed-44 ( 16K) [CUDA0 ]: norm-44 ( 16K) [CUDA0 ] blk.44.attn_q_norm.w ( 1K) [CUDA0 ]
node #1594 ( ROPE): Qcur-44 ( 16K) [CUDA0 ]: Qcur_normed-44 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1595 ( MUL_MAT): Kcur-44 ( 8K) [CUDA0 ]: blk.44.attn_k.weight ( 4M) [CUDA0 ] attn_norm-44 ( 15K) [CUDA0 ]
node #1597 ( RMS_NORM): norm-44 ( 8K) [CUDA0 ]: Kcur-44 (reshaped) ( 8K) [CUDA0 ]
node #1598 ( MUL): Kcur_normed-44 ( 8K) [CUDA0 ]: norm-44 ( 8K) [CUDA0 ] blk.44.attn_k_norm.w ( 1K) [CUDA0 ]
node #1599 ( ROPE): Kcur-44 ( 8K) [CUDA0 ]: Kcur_normed-44 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1600 ( MUL_MAT): Vcur-44 ( 8K) [CUDA0 ]: blk.44.attn_v.weight ( 6M) [CUDA0 ] attn_norm-44 ( 15K) [CUDA0 ]
node #1602 ( CPY): k_cache_view-44 (cop ( 2K) [CUDA0 ]: Kcur-44 ( 8K) [CUDA0 ] k_cache_view-44 ( 2K) [CUDA0 ]
node #1604 ( CPY): v_cache_view-44 (cop ( 2K) [CUDA0 ]: Vcur-44 ( 8K) [CUDA0 ] v_cache_view-44 ( 2K) [CUDA0 ]

SPLIT #90: CPU # 3 inputs: [q-44 ( 16K)] [k-44 ( 544K)] [v-44 ( 544K)]

node #1608 (FLASH_ATTN): node_1608 ( 16K) [ CPU ]: CPU#q-44#0 ( 16K) [ NULL ] CPU#k-44#0 ( 544K) [ NULL ] CPU#v-44#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #91: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1610 ( MUL_MAT): kqv_out-44 ( 15K) [CUDA0 ]: blk.44.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1611 ( RMS_NORM): norm-44 ( 15K) [CUDA0 ]: kqv_out-44 ( 15K) [CUDA0 ]
node #1612 ( MUL): attn_post_norm-44 ( 15K) [CUDA0 ]: norm-44 ( 15K) [CUDA0 ] blk.44.post_attentio ( 15K) [CUDA0 ]
node #1613 ( ADD): sa_out-44 ( 15K) [CUDA0 ]: attn_post_norm-44 ( 15K) [CUDA0 ] l_out-43 ( 15K) [CUDA0 ]
node #1614 ( RMS_NORM): norm-44 ( 15K) [CUDA0 ]: sa_out-44 ( 15K) [CUDA0 ]
node #1615 ( MUL): ffn_norm-44 ( 15K) [CUDA0 ]: norm-44 ( 15K) [CUDA0 ] blk.44.ffn_norm.weig ( 15K) [CUDA0 ]
node #1616 ( MUL_MAT): ffn_gate-44 ( 60K) [CUDA0 ]: blk.44.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-44 ( 15K) [CUDA0 ]
node #1617 ( UNARY): ffn_gelu-44 ( 60K) [CUDA0 ]: ffn_gate-44 ( 60K) [CUDA0 ]
node #1618 ( MUL_MAT): ffn_up-44 ( 60K) [CUDA0 ]: blk.44.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-44 ( 15K) [CUDA0 ]
node #1619 ( MUL): ffn_gate_par-44 ( 60K) [CUDA0 ]: ffn_gelu-44 ( 60K) [CUDA0 ] ffn_up-44 ( 60K) [CUDA0 ]
node #1620 ( MUL_MAT): ffn_out-44 ( 15K) [CUDA0 ]: blk.44.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-44 ( 60K) [CUDA0 ]
node #1621 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-44 ( 15K) [CUDA0 ]
node #1622 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.44.post_ffw_norm ( 15K) [CUDA0 ]
node #1623 ( ADD): l_out-44 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-44 ( 15K) [CUDA0 ]
node #1624 ( RMS_NORM): norm-45 ( 15K) [CUDA0 ]: l_out-44 ( 15K) [CUDA0 ]
node #1625 ( MUL): attn_norm-45 ( 15K) [CUDA0 ]: norm-45 ( 15K) [CUDA0 ] blk.45.attn_norm.wei ( 15K) [CUDA0 ]
node #1626 ( MUL_MAT): Qcur-45 ( 16K) [CUDA0 ]: blk.45.attn_q.weight ( 8M) [CUDA0 ] attn_norm-45 ( 15K) [CUDA0 ]
node #1628 ( RMS_NORM): norm-45 ( 16K) [CUDA0 ]: Qcur-45 (reshaped) ( 16K) [CUDA0 ]
node #1629 ( MUL): Qcur_normed-45 ( 16K) [CUDA0 ]: norm-45 ( 16K) [CUDA0 ] blk.45.attn_q_norm.w ( 1K) [CUDA0 ]
node #1630 ( ROPE): Qcur-45 ( 16K) [CUDA0 ]: Qcur_normed-45 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1631 ( MUL_MAT): Kcur-45 ( 8K) [CUDA0 ]: blk.45.attn_k.weight ( 4M) [CUDA0 ] attn_norm-45 ( 15K) [CUDA0 ]
node #1633 ( RMS_NORM): norm-45 ( 8K) [CUDA0 ]: Kcur-45 (reshaped) ( 8K) [CUDA0 ]
node #1634 ( MUL): Kcur_normed-45 ( 8K) [CUDA0 ]: norm-45 ( 8K) [CUDA0 ] blk.45.attn_k_norm.w ( 1K) [CUDA0 ]
node #1635 ( ROPE): Kcur-45 ( 8K) [CUDA0 ]: Kcur_normed-45 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1636 ( MUL_MAT): Vcur-45 ( 8K) [CUDA0 ]: blk.45.attn_v.weight ( 6M) [CUDA0 ] attn_norm-45 ( 15K) [CUDA0 ]
node #1638 ( CPY): k_cache_view-45 (cop ( 2K) [CUDA0 ]: Kcur-45 ( 8K) [CUDA0 ] k_cache_view-45 ( 2K) [CUDA0 ]
node #1640 ( CPY): v_cache_view-45 (cop ( 2K) [CUDA0 ]: Vcur-45 ( 8K) [CUDA0 ] v_cache_view-45 ( 2K) [CUDA0 ]

SPLIT #92: CPU # 3 inputs: [q-45 ( 16K)] [k-45 ( 544K)] [v-45 ( 544K)]

node #1644 (FLASH_ATTN): node_1644 ( 16K) [ CPU ]: CPU#q-45#0 ( 16K) [ NULL ] CPU#k-45#0 ( 544K) [ NULL ] CPU#v-45#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #93: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1646 ( MUL_MAT): kqv_out-45 ( 15K) [CUDA0 ]: blk.45.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1647 ( RMS_NORM): norm-45 ( 15K) [CUDA0 ]: kqv_out-45 ( 15K) [CUDA0 ]
node #1648 ( MUL): attn_post_norm-45 ( 15K) [CUDA0 ]: norm-45 ( 15K) [CUDA0 ] blk.45.post_attentio ( 15K) [CUDA0 ]
node #1649 ( ADD): sa_out-45 ( 15K) [CUDA0 ]: attn_post_norm-45 ( 15K) [CUDA0 ] l_out-44 ( 15K) [CUDA0 ]
node #1650 ( RMS_NORM): norm-45 ( 15K) [CUDA0 ]: sa_out-45 ( 15K) [CUDA0 ]
node #1651 ( MUL): ffn_norm-45 ( 15K) [CUDA0 ]: norm-45 ( 15K) [CUDA0 ] blk.45.ffn_norm.weig ( 15K) [CUDA0 ]
node #1652 ( MUL_MAT): ffn_gate-45 ( 60K) [CUDA0 ]: blk.45.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-45 ( 15K) [CUDA0 ]
node #1653 ( UNARY): ffn_gelu-45 ( 60K) [CUDA0 ]: ffn_gate-45 ( 60K) [CUDA0 ]
node #1654 ( MUL_MAT): ffn_up-45 ( 60K) [CUDA0 ]: blk.45.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-45 ( 15K) [CUDA0 ]
node #1655 ( MUL): ffn_gate_par-45 ( 60K) [CUDA0 ]: ffn_gelu-45 ( 60K) [CUDA0 ] ffn_up-45 ( 60K) [CUDA0 ]
node #1656 ( MUL_MAT): ffn_out-45 ( 15K) [CUDA0 ]: blk.45.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-45 ( 60K) [CUDA0 ]
node #1657 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-45 ( 15K) [CUDA0 ]
node #1658 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.45.post_ffw_norm ( 15K) [CUDA0 ]
node #1659 ( ADD): l_out-45 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-45 ( 15K) [CUDA0 ]
node #1660 ( RMS_NORM): norm-46 ( 15K) [CUDA0 ]: l_out-45 ( 15K) [CUDA0 ]
node #1661 ( MUL): attn_norm-46 ( 15K) [CUDA0 ]: norm-46 ( 15K) [CUDA0 ] blk.46.attn_norm.wei ( 15K) [CUDA0 ]
node #1662 ( MUL_MAT): Qcur-46 ( 16K) [CUDA0 ]: blk.46.attn_q.weight ( 8M) [CUDA0 ] attn_norm-46 ( 15K) [CUDA0 ]
node #1664 ( RMS_NORM): norm-46 ( 16K) [CUDA0 ]: Qcur-46 (reshaped) ( 16K) [CUDA0 ]
node #1665 ( MUL): Qcur_normed-46 ( 16K) [CUDA0 ]: norm-46 ( 16K) [CUDA0 ] blk.46.attn_q_norm.w ( 1K) [CUDA0 ]
node #1666 ( ROPE): Qcur-46 ( 16K) [CUDA0 ]: Qcur_normed-46 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1667 ( MUL_MAT): Kcur-46 ( 8K) [CUDA0 ]: blk.46.attn_k.weight ( 4M) [CUDA0 ] attn_norm-46 ( 15K) [CUDA0 ]
node #1669 ( RMS_NORM): norm-46 ( 8K) [CUDA0 ]: Kcur-46 (reshaped) ( 8K) [CUDA0 ]
node #1670 ( MUL): Kcur_normed-46 ( 8K) [CUDA0 ]: norm-46 ( 8K) [CUDA0 ] blk.46.attn_k_norm.w ( 1K) [CUDA0 ]
node #1671 ( ROPE): Kcur-46 ( 8K) [CUDA0 ]: Kcur_normed-46 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1672 ( MUL_MAT): Vcur-46 ( 8K) [CUDA0 ]: blk.46.attn_v.weight ( 6M) [CUDA0 ] attn_norm-46 ( 15K) [CUDA0 ]
node #1674 ( CPY): k_cache_view-46 (cop ( 2K) [CUDA0 ]: Kcur-46 ( 8K) [CUDA0 ] k_cache_view-46 ( 2K) [CUDA0 ]
node #1676 ( CPY): v_cache_view-46 (cop ( 2K) [CUDA0 ]: Vcur-46 ( 8K) [CUDA0 ] v_cache_view-46 ( 2K) [CUDA0 ]

SPLIT #94: CPU # 3 inputs: [q-46 ( 16K)] [k-46 ( 544K)] [v-46 ( 544K)]

node #1680 (FLASH_ATTN): node_1680 ( 16K) [ CPU ]: CPU#q-46#0 ( 16K) [ NULL ] CPU#k-46#0 ( 544K) [ NULL ] CPU#v-46#0 ( 544K) [ NULL ] CPU#KQ_mask_swa (cop ( 32K) [ NULL ]

SPLIT #95: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

node #1682 ( MUL_MAT): kqv_out-46 ( 15K) [CUDA0 ]: blk.46.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1683 ( RMS_NORM): norm-46 ( 15K) [CUDA0 ]: kqv_out-46 ( 15K) [CUDA0 ]
node #1684 ( MUL): attn_post_norm-46 ( 15K) [CUDA0 ]: norm-46 ( 15K) [CUDA0 ] blk.46.post_attentio ( 15K) [CUDA0 ]
node #1685 ( ADD): sa_out-46 ( 15K) [CUDA0 ]: attn_post_norm-46 ( 15K) [CUDA0 ] l_out-45 ( 15K) [CUDA0 ]
node #1686 ( RMS_NORM): norm-46 ( 15K) [CUDA0 ]: sa_out-46 ( 15K) [CUDA0 ]
node #1687 ( MUL): ffn_norm-46 ( 15K) [CUDA0 ]: norm-46 ( 15K) [CUDA0 ] blk.46.ffn_norm.weig ( 15K) [CUDA0 ]
node #1688 ( MUL_MAT): ffn_gate-46 ( 60K) [CUDA0 ]: blk.46.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-46 ( 15K) [CUDA0 ]
node #1689 ( UNARY): ffn_gelu-46 ( 60K) [CUDA0 ]: ffn_gate-46 ( 60K) [CUDA0 ]
node #1690 ( MUL_MAT): ffn_up-46 ( 60K) [CUDA0 ]: blk.46.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-46 ( 15K) [CUDA0 ]
node #1691 ( MUL): ffn_gate_par-46 ( 60K) [CUDA0 ]: ffn_gelu-46 ( 60K) [CUDA0 ] ffn_up-46 ( 60K) [CUDA0 ]
node #1692 ( MUL_MAT): ffn_out-46 ( 15K) [CUDA0 ]: blk.46.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-46 ( 60K) [CUDA0 ]
node #1693 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-46 ( 15K) [CUDA0 ]
node #1694 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.46.post_ffw_norm ( 15K) [CUDA0 ]
node #1695 ( ADD): l_out-46 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-46 ( 15K) [CUDA0 ]
node #1696 ( RMS_NORM): norm-47 ( 15K) [CUDA0 ]: l_out-46 ( 15K) [CUDA0 ]
node #1697 ( MUL): attn_norm-47 ( 15K) [CUDA0 ]: norm-47 ( 15K) [CUDA0 ] blk.47.attn_norm.wei ( 15K) [CUDA0 ]
node #1698 ( MUL_MAT): Qcur-47 ( 16K) [CUDA0 ]: blk.47.attn_q.weight ( 8M) [CUDA0 ] attn_norm-47 ( 15K) [CUDA0 ]
node #1700 ( RMS_NORM): norm-47 ( 16K) [CUDA0 ]: Qcur-47 (reshaped) ( 16K) [CUDA0 ]
node #1701 ( MUL): Qcur_normed-47 ( 16K) [CUDA0 ]: norm-47 ( 16K) [CUDA0 ] blk.47.attn_q_norm.w ( 1K) [CUDA0 ]
node #1702 ( ROPE): Qcur-47 ( 16K) [CUDA0 ]: Qcur_normed-47 ( 16K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1703 ( MUL_MAT): Kcur-47 ( 8K) [CUDA0 ]: blk.47.attn_k.weight ( 4M) [CUDA0 ] attn_norm-47 ( 15K) [CUDA0 ]
node #1705 ( RMS_NORM): norm-47 ( 8K) [CUDA0 ]: Kcur-47 (reshaped) ( 8K) [CUDA0 ]
node #1706 ( MUL): Kcur_normed-47 ( 8K) [CUDA0 ]: norm-47 ( 8K) [CUDA0 ] blk.47.attn_k_norm.w ( 1K) [CUDA0 ]
node #1707 ( ROPE): Kcur-47 ( 8K) [CUDA0 ]: Kcur_normed-47 ( 8K) [CUDA0 ] CUDA0#inp_pos#0 ( 0K) [ NULL ]
node #1708 ( MUL_MAT): Vcur-47 ( 8K) [CUDA0 ]: blk.47.attn_v.weight ( 6M) [CUDA0 ] attn_norm-47 ( 15K) [CUDA0 ]
node #1710 ( CPY): k_cache_view-47 (cop ( 2K) [CUDA0 ]: Kcur-47 ( 8K) [CUDA0 ] k_cache_view-47 ( 2K) [CUDA0 ]
node #1712 ( CPY): v_cache_view-47 (cop ( 2K) [CUDA0 ]: Vcur-47 ( 8K) [CUDA0 ] v_cache_view-47 ( 2K) [CUDA0 ]

SPLIT #96: CPU # 3 inputs: [q-47 ( 16K)] [k-47 ( 544K)] [v-47 ( 544K)]

node #1716 (FLASH_ATTN): node_1716 ( 16K) [ CPU ]: CPU#q-47#0 ( 16K) [ NULL ] CPU#k-47#0 ( 544K) [ NULL ] CPU#v-47#0 ( 544K) [ NULL ] CPU#KQ_mask (copy)#0 ( 32K) [ NULL ]

SPLIT #97: CUDA0 # 2 inputs: [ (reshaped) ( 16K)] [inp_out_ids ( 0K)]

node #1718 ( MUL_MAT): kqv_out-47 ( 15K) [CUDA0 ]: blk.47.attn_output.w ( 8M) [CUDA0 ] CUDA0# (reshaped)#0 ( 16K) [ NULL ]
node #1719 ( RMS_NORM): norm-47 ( 15K) [CUDA0 ]: kqv_out-47 ( 15K) [CUDA0 ]
node #1720 ( MUL): attn_post_norm-47 ( 15K) [CUDA0 ]: norm-47 ( 15K) [CUDA0 ] blk.47.post_attentio ( 15K) [CUDA0 ]
node #1721 ( GET_ROWS): node_1721 ( 15K) [CUDA0 ]: attn_post_norm-47 ( 15K) [CUDA0 ] CUDA0#inp_out_ids#0 ( 0K) [ NULL ]
node #1722 ( GET_ROWS): node_1722 ( 15K) [CUDA0 ]: l_out-46 ( 15K) [CUDA0 ] CUDA0#inp_out_ids#0 ( 0K) [ NULL ]
node #1723 ( ADD): sa_out-47 ( 15K) [CUDA0 ]: node_1721 ( 15K) [CUDA0 ] node_1722 ( 15K) [CUDA0 ]
node #1724 ( RMS_NORM): norm-47 ( 15K) [CUDA0 ]: sa_out-47 ( 15K) [CUDA0 ]
node #1725 ( MUL): ffn_norm-47 ( 15K) [CUDA0 ]: norm-47 ( 15K) [CUDA0 ] blk.47.ffn_norm.weig ( 15K) [CUDA0 ]
node #1726 ( MUL_MAT): ffn_gate-47 ( 60K) [CUDA0 ]: blk.47.ffn_gate.weig ( 31M) [CUDA0 ] ffn_norm-47 ( 15K) [CUDA0 ]
node #1727 ( UNARY): ffn_gelu-47 ( 60K) [CUDA0 ]: ffn_gate-47 ( 60K) [CUDA0 ]
node #1728 ( MUL_MAT): ffn_up-47 ( 60K) [CUDA0 ]: blk.47.ffn_up.weight ( 31M) [CUDA0 ] ffn_norm-47 ( 15K) [CUDA0 ]
node #1729 ( MUL): ffn_gate_par-47 ( 60K) [CUDA0 ]: ffn_gelu-47 ( 60K) [CUDA0 ] ffn_up-47 ( 60K) [CUDA0 ]
node #1730 ( MUL_MAT): ffn_out-47 ( 15K) [CUDA0 ]: blk.47.ffn_down.weig ( 46M) [CUDA0 ] ffn_gate_par-47 ( 60K) [CUDA0 ]
node #1731 ( RMS_NORM): norm ( 15K) [CUDA0 ]: ffn_out-47 ( 15K) [CUDA0 ]
node #1732 ( MUL): ffn_post_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] blk.47.post_ffw_norm ( 15K) [CUDA0 ]
node #1733 ( ADD): l_out-47 ( 15K) [CUDA0 ]: ffn_post_norm ( 15K) [CUDA0 ] sa_out-47 ( 15K) [CUDA0 ]
node #1734 ( RMS_NORM): norm ( 15K) [CUDA0 ]: l_out-47 ( 15K) [CUDA0 ]
node #1735 ( MUL): result_norm ( 15K) [CUDA0 ]: norm ( 15K) [CUDA0 ] output_norm.weight ( 15K) [CUDA0 ]
node #1736 ( MUL_MAT): result_output ( 1M) [CUDA0 ]: token_embd.weight ( 787M) [CUDA0 ] result_norm ( 15K) [CUDA0 ]
srv send: sending result for task id = 7
srv send: task id = 7 pushed to result queue
slot process_toke: id 0 | task 7 | stopped by EOS
slot process_toke: id 0 | task 7 | n_decoded = 22, n_remaining = -1, next token: 106 ''
slot release: id 0 | task 7 | stop processing: n_past = 38, truncated = 0
slot print_timing: id 0 | task 7 |
prompt eval time = 75.86 ms / 13 tokens ( 5.84 ms per token, 171.38 tokens per second)
eval time = 945.03 ms / 22 tokens ( 42.96 ms per token, 23.28 tokens per second)
total time = 1020.88 ms / 35 tokens
srv send: sending result for task id = 7
srv send: task id = 7 pushed to result queue
srv update_slots: run slots completed
que start_loop: waiting for new tasks
que start_loop: processing new tasks
que start_loop: processing task, id = 29
que start_loop: update slots
srv update_slots: all slots are idle
que start_loop: waiting for new tasks
data stream, to_send: data: {"choices":[{"finish_reason":null,"index":0,"delta":{"content":""}}],"created":1741785946,"id":"chatcmpl-9Qb2GeBPXWugyHyvaWal9u90NNwaReFo","model":"gpt-3.5-turbo","system_fingerprint":"b0-unknown","object":"chat.completion.chunk"}

data stream, to_send: data: {"choices":[{"finish_reason":"stop","index":0,"delta":{}}],"created":1741785946,"id":"chatcmpl-9Qb2GeBPXWugyHyvaWal9u90NNwaReFo","model":"gpt-3.5-turbo","system_fingerprint":"b0-unknown","object":"chat.completion.chunk","usage":{"completion_tokens":22,"prompt_tokens":17,"total_tokens":39},"timings":{"prompt_n":13,"prompt_ms":75.857,"prompt_per_token_ms":5.835153846153846,"prompt_per_second":171.37508733538104,"predicted_n":22,"predicted_ms":945.028,"predicted_per_token_ms":42.95581818181818,"predicted_per_second":23.279733510541487}}

srv log_server_r: request: POST /v1/chat/completions 127.0.0.1 200
srv log_server_r: request: {"messages":[{"role":"system","content":"You are a helpful assistant."},{"role":"user","content":"Hello"}],"stream":true,"cache_prompt":true,"samplers":"edkypmxt","temperature":0.8,"dynatemp_range":0,"dynatemp_exponent":1,"top_k":40,"top_p":0.95,"min_p":0.05,"typical_p":1,"xtc_probability":0,"xtc_threshold":0.1,"repeat_last_n":64,"repeat_penalty":1,"presence_penalty":0,"frequency_penalty":0,"dry_multiplier":0,"dry_base":1.75,"dry_allowed_length":2,"dry_penalty_last_n":-1,"max_tokens":-1,"timings_per_token":false}
srv log_server_r: response:
srv remove_waiti: remove task 7 from waiting list. current waiting = 1 (before remove)

slaren · 2025-03-12T13:32:21Z

So it is the flash attention. This is probably because this head size (256) is only supported with F16. Not sure if this is because it is not commonly used, or there is some performance issue that makes it unusable, @JohannesGaessler should know more. You should still be able to use K quantization with flash attention disabled.

Bearsaerker · 2025-03-12T13:37:15Z

Without the flash attention flag it does not load at all unfortunately. The command:

./bin/llama-server -m '/home/luis/Downloads/gemma-3-12b-it-Q4_K_M.gguf' --n-gpu-layers -1 --batch_size 1024 --cache-type-k q8_0 --cache-type-v q8_0 -c 8000 --port 7777 -t 8 -ngl 99 -v

the logs:
...
load: control token: 259067 '' is not marked as EOG
load: control token: 261192 '' is not marked as EOG
load: control token: 261582 '' is not marked as EOG
load: control token: 260252 '' is not marked as EOG
load: control token: 258651 '' is not marked as EOG
load: control token: 259827 '' is not marked as EOG
load: control token: 258030 '' is not marked as EOG
load: control token: 259274 '' is not marked as EOG
load: control token: 260600 '' is not marked as EOG
load: control token: 259848 '' is not marked as EOG
load: control token: 261065 '' is not marked as EOG
load: control token: 260213 '' is not marked as EOG
load: control token: 259188 '' is not marked as EOG
load: control token: 257859 '' is not marked as EOG
load: control token: 259114 '' is not marked as EOG
load: control token: 261098 '' is not marked as EOG
load: control token: 259756 '' is not marked as EOG
load: control token: 257649 '' is not marked as EOG
load: control token: 256694 '' is not marked as EOG
load: control token: 258299 '' is not marked as EOG
load: control token: 259890 '' is not marked as EOG
load: control token: 261474 '' is not marked as EOG
load: control token: 258611 '' is not marked as EOG
load: control token: 258900 '' is not marked as EOG
load: control token: 258655 '' is not marked as EOG
load: control token: 257528 '' is not marked as EOG
load: control token: 260064 '' is not marked as EOG
load: control token: 257706 '' is not marked as EOG
load: control token: 256635 '' is not marked as EOG
load: control token: 256515 '' is not marked as EOG
load: control token: 258337 '' is not marked as EOG
load: control token: 258105 '' is not marked as EOG
load: control token: 260748 '' is not marked as EOG
load: control token: 261439 '' is not marked as EOG
load: control token: 260788 '' is not marked as EOG
load: control token: 256904 '' is not marked as EOG
load: control token: 261931 '' is not marked as EOG
load: control token: 257320 '' is not marked as EOG
load: control token: 256602 '' is not marked as EOG
load: control token: 256144 '' is not marked as EOG
load: control token: 257785 '' is not marked as EOG
load: control token: 261560 '' is not marked as EOG
load: control token: 260631 '' is not marked as EOG
load: control token: 258591 '' is not marked as EOG
load: control token: 257519 '' is not marked as EOG
load: control token: 259283 '' is not marked as EOG
load: control token: 256126 '' is not marked as EOG
load: control token: 257728 '' is not marked as EOG
load: control token: 260871 '' is not marked as EOG
load: control token: 260829 '' is not marked as EOG
load: control token: 260996 '' is not marked as EOG
load: control token: 259635 '' is not marked as EOG
load: control token: 257056 '' is not marked as EOG
load: control token: 256416 '' is not marked as EOG
load: control token: 259863 '' is not marked as EOG
load: control token: 258343 '' is not marked as EOG
load: control token: 257857 '' is not marked as EOG
load: control token: 261383 '' is not marked as EOG
load: control token: 261730 '' is not marked as EOG
load: control token: 261228 '' is not marked as EOG
load: control token: 258397 '' is not marked as EOG
load: control token: 259751 '' is not marked as EOG
load: control token: 257544 '' is not marked as EOG
load: control token: 256615 '' is not marked as EOG
load: control token: 259972 '' is not marked as EOG
load: control token: 261915 '' is not marked as EOG
load: control token: 258188 '' is not marked as EOG
load: control token: 258544 '' is not marked as EOG
load: control token: 256511 '' is not marked as EOG
load: control token: 258958 '' is not marked as EOG
load: control token: 261422 '' is not marked as EOG
load: control token: 258098 '' is not marked as EOG
load: control token: 262025 '' is not marked as EOG
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 6414
load: token to piece cache size = 1.9446 MB
print_info: arch = gemma3
print_info: vocab_only = 0
print_info: n_ctx_train = 131072
print_info: n_embd = 3840
print_info: n_layer = 48
print_info: n_head = 16
print_info: n_head_kv = 8
print_info: n_rot = 256
print_info: n_swa = 1024
print_info: n_embd_head_k = 256
print_info: n_embd_head_v = 256
print_info: n_gqa = 2
print_info: n_embd_k_gqa = 2048
print_info: n_embd_v_gqa = 2048
print_info: f_norm_eps = 0.0e+00
print_info: f_norm_rms_eps = 1.0e-06
print_info: f_clamp_kqv = 0.0e+00
print_info: f_max_alibi_bias = 0.0e+00
print_info: f_logit_scale = 0.0e+00
print_info: f_attn_scale = 6.2e-02
print_info: n_ff = 15360
print_info: n_expert = 0
print_info: n_expert_used = 0
print_info: causal attn = 1
print_info: pooling type = 0
print_info: rope type = 2
print_info: rope scaling = linear
print_info: freq_base_train = 1000000.0
print_info: freq_scale_train = 0.125
print_info: n_ctx_orig_yarn = 131072
print_info: rope_finetuned = unknown
print_info: ssm_d_conv = 0
print_info: ssm_d_inner = 0
print_info: ssm_d_state = 0
print_info: ssm_dt_rank = 0
print_info: ssm_dt_b_c_rms = 0
print_info: model type = 12B
print_info: model params = 11.77 B
print_info: general.name = Gemma 3 12b It
print_info: vocab type = SPM
print_info: n_vocab = 262144
print_info: n_merges = 0
print_info: BOS token = 2 ''
print_info: EOS token = 1 ''
print_info: EOT token = 106 '<end_of_turn>'
print_info: UNK token = 3 ''
print_info: PAD token = 0 ''
print_info: LF token = 248 '<0x0A>'
print_info: EOG token = 1 ''
print_info: EOG token = 106 '<end_of_turn>'
print_info: max token length = 48
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: layer 0 assigned to device CUDA0
load_tensors: layer 1 assigned to device CUDA0
load_tensors: layer 2 assigned to device CUDA0
load_tensors: layer 3 assigned to device CUDA0
load_tensors: layer 4 assigned to device CUDA0
load_tensors: layer 5 assigned to device CUDA0
load_tensors: layer 6 assigned to device CUDA0
load_tensors: layer 7 assigned to device CUDA0
load_tensors: layer 8 assigned to device CUDA0
load_tensors: layer 9 assigned to device CUDA0
load_tensors: layer 10 assigned to device CUDA0
load_tensors: layer 11 assigned to device CUDA0
load_tensors: layer 12 assigned to device CUDA0
load_tensors: layer 13 assigned to device CUDA0
load_tensors: layer 14 assigned to device CUDA0
load_tensors: layer 15 assigned to device CUDA0
load_tensors: layer 16 assigned to device CUDA0
load_tensors: layer 17 assigned to device CUDA0
load_tensors: layer 18 assigned to device CUDA0
load_tensors: layer 19 assigned to device CUDA0
load_tensors: layer 20 assigned to device CUDA0
load_tensors: layer 21 assigned to device CUDA0
load_tensors: layer 22 assigned to device CUDA0
load_tensors: layer 23 assigned to device CUDA0
load_tensors: layer 24 assigned to device CUDA0
load_tensors: layer 25 assigned to device CUDA0
load_tensors: layer 26 assigned to device CUDA0
load_tensors: layer 27 assigned to device CUDA0
load_tensors: layer 28 assigned to device CUDA0
load_tensors: layer 29 assigned to device CUDA0
load_tensors: layer 30 assigned to device CUDA0
load_tensors: layer 31 assigned to device CUDA0
load_tensors: layer 32 assigned to device CUDA0
load_tensors: layer 33 assigned to device CUDA0
load_tensors: layer 34 assigned to device CUDA0
load_tensors: layer 35 assigned to device CUDA0
load_tensors: layer 36 assigned to device CUDA0
load_tensors: layer 37 assigned to device CUDA0
load_tensors: layer 38 assigned to device CUDA0
load_tensors: layer 39 assigned to device CUDA0
load_tensors: layer 40 assigned to device CUDA0
load_tensors: layer 41 assigned to device CUDA0
load_tensors: layer 42 assigned to device CUDA0
load_tensors: layer 43 assigned to device CUDA0
load_tensors: layer 44 assigned to device CUDA0
load_tensors: layer 45 assigned to device CUDA0
load_tensors: layer 46 assigned to device CUDA0
load_tensors: layer 47 assigned to device CUDA0
load_tensors: layer 48 assigned to device CUDA0
load_tensors: tensor 'token_embd.weight' (q6_K) (and 0 others) cannot be used with preferred buffer type CPU_AARCH64, using CPU instead
load_tensors: offloading 48 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 49/49 layers to GPU
load_tensors: CUDA0 model buffer size = 6956.18 MiB
load_tensors: CPU_Mapped model buffer size = 787.50 MiB
.................................................................................
llama_init_from_model: V cache quantization requires flash_attn
common_init_from_params: failed to create context with model '/home/luis/Downloads/gemma-3-12b-it-Q4_K_M.gguf'
srv load_model: failed to load model, '/home/luis/Downloads/gemma-3-12b-it-Q4_K_M.gguf'
srv operator(): operator(): cleaning up before exit...
terminate called without an active exception
main: exiting due to model loading error

slaren · 2025-03-12T13:38:50Z

Without flash attn you can only quantize K, but not V. You need to remove the --cache-type-v q8_0 option.

BVEsun · 2025-03-13T01:13:18Z

I had encounter the same problem, seems it auto kv offload even without -nkvo option

MMaturax · 2025-03-13T07:23:36Z

Models with the same parameter count run significantly faster, and even models with a larger parameter count perform better than Gemma 3.

Related issue: GitHub Issue #9701

ag2s20150909 · 2025-03-13T13:21:23Z

try -fa -ctk q4_0 -ctv q4_0

Bearsaerker · 2025-03-13T14:19:25Z

changes nothing, same output

icsy7867 · 2025-03-15T01:28:18Z

I'm seeing the same on my older P6000.

JNLLM · 2025-03-15T11:12:47Z

Same case, while I can run 12b models easily, gemma3 12b gets its cache offloaded. And not having v cache quantized is not an option for low vram situations. If I use --flash-attn -ctk q4_0 -ctv q4_0 the prompt processing is done in CPU. If I user --flash-attn -ctk q4_0 the promt processing is offloaded but the vram consumption skyrockets out of the gpu capacity and that is slow as molases.

Changing topic a bit, I am really grateful for all the comunity efforts on llama.cpp.

BahamutRU · 2025-03-15T12:33:58Z

Inference speed double slow, when use q8_0 cache.

16 token/sec unquantized vs 8 token/sec with q8_0 kv cache.

In mistral nemo this ~5% slower.

BahamutRU · 2025-03-16T13:43:03Z

Up! Very sad bug: fat context, but quantized is kills inference speed. =(

JohannesGaessler · 2025-03-16T13:53:27Z

So it is the flash attention. This is probably because this head size (256) is only supported with F16. Not sure if this is because it is not commonly used, or there is some performance issue that makes it unusable, @JohannesGaessler should know more. You should still be able to use K quantization with flash attention disabled.

The problem is register pressure. Head size 256 needs more registers than head size 128 and a quantized KV cache also needs more registers than an FP16 KV cache. If you combine the two the current kernel simply runs out of registers and the performance is effectively unusable which is why the CUDA backend does not support it. The code would need to be specifically rewritten for that use case to make it usable.

BahamutRU · 2025-03-16T16:20:36Z

The problem is register pressure. Head size 256 needs more registers than head size 128 and a quantized KV cache also needs more registers than an FP16 KV cache. If you combine the two the current kernel simply runs out of registers and the performance is effectively unusable which is why the CUDA backend does not support it. The code would need to be specifically rewritten for that use case to make it usable.

Thx for your answer!

Bearsaerker added the bug-unconfirmed label Mar 12, 2025

icsy7867 mentioned this issue Mar 16, 2025

Gemma 3 KV cache quantization issue LostRuins/koboldcpp#1423

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Eval bug: Gemma 3 extremly slow prompt processing when using quantized kv cache. #12352

Eval bug: Gemma 3 extremly slow prompt processing when using quantized kv cache. #12352

Bearsaerker commented Mar 12, 2025

Andybui1012 commented Mar 12, 2025

szghds commented Mar 12, 2025

slaren commented Mar 12, 2025

Bearsaerker commented Mar 12, 2025

slaren commented Mar 12, 2025

Bearsaerker commented Mar 12, 2025

slaren commented Mar 12, 2025 •

edited

Loading

Bearsaerker commented Mar 12, 2025

slaren commented Mar 12, 2025

BVEsun commented Mar 13, 2025

MMaturax commented Mar 13, 2025

ag2s20150909 commented Mar 13, 2025

Bearsaerker commented Mar 13, 2025

icsy7867 commented Mar 15, 2025

JNLLM commented Mar 15, 2025

BahamutRU commented Mar 15, 2025

BahamutRU commented Mar 16, 2025

JohannesGaessler commented Mar 16, 2025

BahamutRU commented Mar 16, 2025

Eval bug: Gemma 3 extremly slow prompt processing when using quantized kv cache. #12352

Eval bug: Gemma 3 extremly slow prompt processing when using quantized kv cache. #12352

Comments

Bearsaerker commented Mar 12, 2025

Name and Version

Operating systems

GGML backends

Hardware

Models

Problem description & steps to reproduce

First Bad Commit

Relevant log output

Andybui1012 commented Mar 12, 2025

szghds commented Mar 12, 2025

slaren commented Mar 12, 2025

Bearsaerker commented Mar 12, 2025

slaren commented Mar 12, 2025

Bearsaerker commented Mar 12, 2025

SPLIT #0: CPU # 0 inputs

SPLIT #1: CUDA0 # 3 inputs: [inp_embd ( 15K)] [inp_pos ( 0K)] [KQ_mask_swa ( 64K)]

SPLIT #2: CPU # 4 inputs: [q-0 ( 16K)] [k-0 ( 544K)] [v-0 ( 544K)] [KQ_mask_swa (copy) ( 32K)]

SPLIT #3: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #4: CPU # 3 inputs: [q-1 ( 16K)] [k-1 ( 544K)] [v-1 ( 544K)]

SPLIT #5: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #6: CPU # 3 inputs: [q-2 ( 16K)] [k-2 ( 544K)] [v-2 ( 544K)]

SPLIT #7: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #8: CPU # 3 inputs: [q-3 ( 16K)] [k-3 ( 544K)] [v-3 ( 544K)]

SPLIT #9: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #10: CPU # 3 inputs: [q-4 ( 16K)] [k-4 ( 544K)] [v-4 ( 544K)]

SPLIT #11: CUDA0 # 2 inputs: [ (reshaped) ( 16K)] [KQ_mask ( 64K)]

SPLIT #12: CPU # 4 inputs: [q-5 ( 16K)] [k-5 ( 544K)] [v-5 ( 544K)] [KQ_mask (copy) ( 32K)]

SPLIT #13: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #14: CPU # 3 inputs: [q-6 ( 16K)] [k-6 ( 544K)] [v-6 ( 544K)]

SPLIT #15: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #16: CPU # 3 inputs: [q-7 ( 16K)] [k-7 ( 544K)] [v-7 ( 544K)]

SPLIT #17: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #18: CPU # 3 inputs: [q-8 ( 16K)] [k-8 ( 544K)] [v-8 ( 544K)]

SPLIT #19: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #20: CPU # 3 inputs: [q-9 ( 16K)] [k-9 ( 544K)] [v-9 ( 544K)]

SPLIT #21: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #22: CPU # 3 inputs: [q-10 ( 16K)] [k-10 ( 544K)] [v-10 ( 544K)]

SPLIT #23: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #24: CPU # 3 inputs: [q-11 ( 16K)] [k-11 ( 544K)] [v-11 ( 544K)]

SPLIT #25: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #26: CPU # 3 inputs: [q-12 ( 16K)] [k-12 ( 544K)] [v-12 ( 544K)]

SPLIT #27: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #28: CPU # 3 inputs: [q-13 ( 16K)] [k-13 ( 544K)] [v-13 ( 544K)]

SPLIT #29: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #30: CPU # 3 inputs: [q-14 ( 16K)] [k-14 ( 544K)] [v-14 ( 544K)]

SPLIT #31: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #32: CPU # 3 inputs: [q-15 ( 16K)] [k-15 ( 544K)] [v-15 ( 544K)]

SPLIT #33: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #34: CPU # 3 inputs: [q-16 ( 16K)] [k-16 ( 544K)] [v-16 ( 544K)]

SPLIT #35: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #36: CPU # 3 inputs: [q-17 ( 16K)] [k-17 ( 544K)] [v-17 ( 544K)]

SPLIT #37: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #38: CPU # 3 inputs: [q-18 ( 16K)] [k-18 ( 544K)] [v-18 ( 544K)]

SPLIT #39: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #40: CPU # 3 inputs: [q-19 ( 16K)] [k-19 ( 544K)] [v-19 ( 544K)]

SPLIT #41: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #42: CPU # 3 inputs: [q-20 ( 16K)] [k-20 ( 544K)] [v-20 ( 544K)]

SPLIT #43: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #44: CPU # 3 inputs: [q-21 ( 16K)] [k-21 ( 544K)] [v-21 ( 544K)]

SPLIT #45: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #46: CPU # 3 inputs: [q-22 ( 16K)] [k-22 ( 544K)] [v-22 ( 544K)]

SPLIT #47: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #48: CPU # 3 inputs: [q-23 ( 16K)] [k-23 ( 544K)] [v-23 ( 544K)]

SPLIT #49: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #50: CPU # 3 inputs: [q-24 ( 16K)] [k-24 ( 544K)] [v-24 ( 544K)]

SPLIT #51: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #52: CPU # 3 inputs: [q-25 ( 16K)] [k-25 ( 544K)] [v-25 ( 544K)]

SPLIT #53: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #54: CPU # 3 inputs: [q-26 ( 16K)] [k-26 ( 544K)] [v-26 ( 544K)]

SPLIT #55: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #56: CPU # 3 inputs: [q-27 ( 16K)] [k-27 ( 544K)] [v-27 ( 544K)]

SPLIT #57: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #58: CPU # 3 inputs: [q-28 ( 16K)] [k-28 ( 544K)] [v-28 ( 544K)]

SPLIT #59: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

SPLIT #60: CPU # 3 inputs: [q-29 ( 16K)] [k-29 ( 544K)] [v-29 ( 544K)]

SPLIT #61: CUDA0 # 1 inputs: [ (reshaped) ( 16K)]

slaren commented Mar 12, 2025 •

edited

Loading