context_with_model: flash_attn = 1 does not work as fast as expected

[hzgdeerHo]

I launched llama_cpp.server as follows:

(base) ubuntu@ly-rq-214-23-49-1:~$ python -m llama_cpp.server --config_file llama_cpp_config.json
llama_model_loader: loaded meta data with 25 key-value pairs and 435 tensors from /home/ubuntu/.cache/huggingface/hub/models--01-ai--Yi-1.5-9B-Chat/models--01-ai--Yi-1.5-9B-Chat-q8_0.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 = llama
llama_model_loader: - kv 1: general.name str = 6afa72fa85c12128e9716fc189b6fc21fe26da83
llama_model_loader: - kv 2: llama.block_count u32 = 48
llama_model_loader: - kv 3: llama.context_length u32 = 4096
llama_model_loader: - kv 4: llama.embedding_length u32 = 4096
llama_model_loader: - kv 5: llama.feed_forward_length u32 = 11008
llama_model_loader: - kv 6: llama.attention.head_count u32 = 32
llama_model_loader: - kv 7: llama.attention.head_count_kv u32 = 4
llama_model_loader: - kv 8: llama.rope.freq_base f32 = 5000000.000000
llama_model_loader: - kv 9: llama.attention.layer_norm_rms_epsilon f32 = 0.000001
llama_model_loader: - kv 10: general.file_type u32 = 7
llama_model_loader: - kv 11: llama.vocab_size u32 = 64000
llama_model_loader: - kv 12: llama.rope.dimension_count u32 = 128
llama_model_loader: - kv 13: tokenizer.ggml.model str = llama
llama_model_loader: - kv 14: tokenizer.ggml.pre str = default
llama_model_loader: - kv 15: tokenizer.ggml.tokens arr[str,64000] = ["", "<|startoftext|>", "<|endof...
llama_model_loader: - kv 16: tokenizer.ggml.scores arr[f32,64000] = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv 17: tokenizer.ggml.token_type arr[i32,64000] = [2, 3, 3, 3, 3, 3, 1, 1, 1, 3, 3, 3, ...
llama_model_loader: - kv 18: tokenizer.ggml.bos_token_id u32 = 1
llama_model_loader: - kv 19: tokenizer.ggml.eos_token_id u32 = 2
llama_model_loader: - kv 20: tokenizer.ggml.padding_token_id u32 = 0
llama_model_loader: - kv 21: tokenizer.ggml.add_bos_token bool = false
llama_model_loader: - kv 22: tokenizer.ggml.add_eos_token bool = false
llama_model_loader: - kv 23: tokenizer.chat_template str = {% if messages[0]['role'] == 'system'...
llama_model_loader: - kv 24: general.quantization_version u32 = 2
llama_model_loader: - type f32: 97 tensors
llama_model_loader: - type q8_0: 338 tensors
llm_load_vocab: mismatch in special tokens definition ( 498/64000 vs 267/64000 ).
llm_load_print_meta: format = GGUF V3 (latest)
llm_load_print_meta: arch = llama
llm_load_print_meta: vocab type = SPM
llm_load_print_meta: n_vocab = 64000
llm_load_print_meta: n_merges = 0
llm_load_print_meta: n_ctx_train = 4096
llm_load_print_meta: n_embd = 4096
llm_load_print_meta: n_head = 32
llm_load_print_meta: n_head_kv = 4
llm_load_print_meta: n_layer = 48
llm_load_print_meta: n_rot = 128
llm_load_print_meta: n_embd_head_k = 128
llm_load_print_meta: n_embd_head_v = 128
llm_load_print_meta: n_gqa = 8
llm_load_print_meta: n_embd_k_gqa = 512
llm_load_print_meta: n_embd_v_gqa = 512
llm_load_print_meta: f_norm_eps = 0.0e+00
llm_load_print_meta: f_norm_rms_eps = 1.0e-06
llm_load_print_meta: f_clamp_kqv = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: f_logit_scale = 0.0e+00
llm_load_print_meta: n_ff = 11008
llm_load_print_meta: n_expert = 0
llm_load_print_meta: n_expert_used = 0
llm_load_print_meta: causal attn = 1
llm_load_print_meta: pooling type = 0
llm_load_print_meta: rope type = 0
llm_load_print_meta: rope scaling = linear
llm_load_print_meta: freq_base_train = 5000000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_yarn_orig_ctx = 4096
llm_load_print_meta: rope_finetuned = unknown
llm_load_print_meta: ssm_d_conv = 0
llm_load_print_meta: ssm_d_inner = 0
llm_load_print_meta: ssm_d_state = 0
llm_load_print_meta: ssm_dt_rank = 0
llm_load_print_meta: model type = 34B
llm_load_print_meta: model ftype = Q8_0
llm_load_print_meta: model params = 8.83 B
llm_load_print_meta: model size = 8.74 GiB (8.50 BPW)
llm_load_print_meta: general.name = 6afa72fa85c12128e9716fc189b6fc21fe26da83
llm_load_print_meta: BOS token = 1 '<|startoftext|>'
llm_load_print_meta: EOS token = 2 '<|endoftext|>'
llm_load_print_meta: UNK token = 0 ''
llm_load_print_meta: PAD token = 0 ''
llm_load_print_meta: LF token = 315 '<0x0A>'
llm_load_print_meta: EOT token = 7 '<|im_end|>'
ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 4 CUDA devices:
Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
Device 1: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
Device 2: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
Device 3: NVIDIA GeForce RTX 4090, compute capability 8.9, VMM: yes
llm_load_tensors: ggml ctx size = 1.11 MiB
llm_load_tensors: offloading 48 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 49/49 layers to GPU
llm_load_tensors: CPU buffer size = 265.62 MiB
llm_load_tensors: CUDA0 buffer size = 2279.47 MiB
llm_load_tensors: CUDA1 buffer size = 2104.13 MiB
llm_load_tensors: CUDA2 buffer size = 2104.13 MiB
llm_load_tensors: CUDA3 buffer size = 2194.43 MiB
.................................................................................................
llama_new_context_with_model: n_ctx = 13568
llama_new_context_with_model: n_batch = 2048
llama_new_context_with_model: n_ubatch = 512
llama_new_context_with_model: flash_attn = 1
llama_new_context_with_model: freq_base = 5000000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init: CUDA0 KV buffer size = 344.50 MiB
llama_kv_cache_init: CUDA1 KV buffer size = 318.00 MiB
llama_kv_cache_init: CUDA2 KV buffer size = 318.00 MiB
llama_kv_cache_init: CUDA3 KV buffer size = 291.50 MiB
llama_new_context_with_model: KV self size = 1272.00 MiB, K (f16): 636.00 MiB, V (f16): 636.00 MiB
llama_new_context_with_model: CUDA_Host output buffer size = 0.26 MiB
llama_new_context_with_model: pipeline parallelism enabled (n_copies=4)
llama_new_context_with_model: CUDA0 compute buffer size = 210.26 MiB
llama_new_context_with_model: CUDA1 compute buffer size = 152.01 MiB
llama_new_context_with_model: CUDA2 compute buffer size = 152.01 MiB
llama_new_context_with_model: CUDA3 compute buffer size = 218.02 MiB
llama_new_context_with_model: CUDA_Host compute buffer size = 114.02 MiB
llama_new_context_with_model: graph nodes = 1351
llama_new_context_with_model: graph splits = 5
AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 |
Model metadata: {'tokenizer.chat_template': "{% if messages[0]['role'] == 'system' %}{% set system_message = messages[0]['content'] %}{% endif %}{% if system_message is defined %}{{ system_message }}{% endif %}{% for message in messages %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|im_start|>user\n' + content + '<|im_end|>\n<|im_start|>assistant\n' }}{% elif message['role'] == 'assistant' %}{{ content + '<|im_end|>' + '\n' }}{% endif %}{% endfor %}", 'tokenizer.ggml.add_eos_token': 'false', 'tokenizer.ggml.padding_token_id': '0', 'tokenizer.ggml.eos_token_id': '2', 'general.quantization_version': '2', 'tokenizer.ggml.model': 'llama', 'general.architecture': 'llama', 'llama.rope.freq_base': '5000000.000000', 'tokenizer.ggml.pre': 'default', 'llama.context_length': '4096', 'general.name': '6afa72fa85c12128e9716fc189b6fc21fe26da83', 'tokenizer.ggml.add_bos_token': 'false', 'llama.embedding_length': '4096', 'llama.feed_forward_length': '11008', 'llama.attention.layer_norm_rms_epsilon': '0.000001', 'tokenizer.ggml.bos_token_id': '1', 'llama.attention.head_count': '32', 'llama.block_count': '48', 'llama.attention.head_count_kv': '4', 'general.file_type': '7', 'llama.vocab_size': '64000', 'llama.rope.dimension_count': '128'}
Available chat formats from metadata: chat_template.default
INFO: Started server process [791036]
INFO: Waiting for application startup.
INFO: Application startup complete.
INFO: Uvicorn running on http://0.0.0.0:8090 (Press CTRL+C to quit)

###Hardware info:

(base) ubuntu@ly-rq-214-23-49-1:~$ nvidia-smi
Tue Jun 4 12:20:52 2024
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.54.15 Driver Version: 550.54.15 CUDA Version: 12.4 |
|-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 4090 On | 00000000:06:00.0 Off | Off |
| 31% 34C P8 16W / 450W | 3348MiB / 24564MiB | 0% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
| 1 NVIDIA GeForce RTX 4090 On | 00000000:07:00.0 Off | Off |
| 30% 32C P8 12W / 450W | 3088MiB / 24564MiB | 0% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
| 2 NVIDIA GeForce RTX 4090 On | 00000000:0D:00.0 Off | Off |
| 31% 32C P8 29W / 450W | 3088MiB / 24564MiB | 0% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
| 3 NVIDIA GeForce RTX 4090 On | 00000000:0E:00.0 Off | Off |
| 30% 35C P8 20W / 450W | 3684MiB / 24564MiB | 0% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 0 N/A N/A 791036 C python 3342MiB |
| 1 N/A N/A 791036 C python 3082MiB |
| 2 N/A N/A 791036 C python 3082MiB |
| 3 N/A N/A 791036 C python 3678MiB |
+-----------------------------------------------------------------------------------------+

####The question: why there is no improved inferring speed after FlashAttention is turned on.

###An acceleration is expected to happen.