Flash Attention not working with NVIDIA Quadro P3200 Pascal Architecture GPU

[countzero]

I am using the server of llama.cpp b2781 on a Windows machine with enabled --flash-attn option.

The Compute Capability of the Quadro P3200 GPU in the machine is 6.1. The installed NVIDIA Driver is at 552.22 and CUDA 12.3.

I am getting the following error message:

ERROR: CUDA kernel flash_attn_ext_f16 has no device code compatible with CUDA arch 610. ggml-cuda.cu was compiled for: 520,610,700

Log

system_info":"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 = 0 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | "}
llama_model_loader: loaded meta data with 24 key-value pairs and 291 tensors from .\vendor\llama.cpp\models\openchat-3.5-0106.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              = llama
llama_model_loader: - kv   1:                               general.name str              = D:\Arbeit\windows_manage_large_langua...
llama_model_loader: - kv   2:                           llama.vocab_size u32              = 32002
llama_model_loader: - kv   3:                       llama.context_length u32              = 8192
llama_model_loader: - kv   4:                     llama.embedding_length u32              = 4096
llama_model_loader: - kv   5:                          llama.block_count u32              = 32
llama_model_loader: - kv   6:                  llama.feed_forward_length u32              = 14336
llama_model_loader: - kv   7:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv   8:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   9:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv  10:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  11:                       llama.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  12:                          general.file_type u32              = 17
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  14:                      tokenizer.ggml.tokens arr[str,32002]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
llama_model_loader: - kv  15:                      tokenizer.ggml.scores arr[f32,32002]   = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  16:                  tokenizer.ggml.token_type arr[i32,32002]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
llama_model_loader: - kv  17:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  18:                tokenizer.ggml.eos_token_id u32              = 32000
llama_model_loader: - kv  19:            tokenizer.ggml.unknown_token_id u32              = 0
llama_model_loader: - kv  20:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  21:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  22:                    tokenizer.chat_template str              = {{ bos_token }}{% for message in mess...
llama_model_loader: - kv  23:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   65 tensors
llama_model_loader: - type q5_K:  193 tensors
llama_model_loader: - type q6_K:   33 tensors
llm_load_vocab: special tokens definition check successful ( 261/32002 ).
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          = 32002
llm_load_print_meta: n_merges         = 0
llm_load_print_meta: n_ctx_train      = 8192
llm_load_print_meta: n_embd           = 4096
llm_load_print_meta: n_head           = 32
llm_load_print_meta: n_head_kv        = 8
llm_load_print_meta: n_layer          = 32
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            = 4
llm_load_print_meta: n_embd_k_gqa     = 1024
llm_load_print_meta: n_embd_v_gqa     = 1024
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-05
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             = 14336
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  = 10000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_yarn_orig_ctx  = 8192
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       = 8B
llm_load_print_meta: model ftype      = Q5_K - Medium
llm_load_print_meta: model params     = 7.24 B
llm_load_print_meta: model size       = 4.78 GiB (5.67 BPW)
llm_load_print_meta: general.name     = D:\Arbeit\windows_manage_large_language_models\source
llm_load_print_meta: BOS token        = 1 '<s>'
llm_load_print_meta: EOS token        = 32000 '<|end_of_turn|>'
llm_load_print_meta: UNK token        = 0 '<unk>'
llm_load_print_meta: LF token         = 13 '<0x0A>'
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 1 CUDA devices:
  Device 0: Quadro P3200 with Max-Q Design, compute capability 6.1, VMM: yes
llm_load_tensors: ggml ctx size =    0.30 MiB
llm_load_tensors: offloading 32 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloaded 33/33 layers to GPU
llm_load_tensors:        CPU buffer size =    85.94 MiB
llm_load_tensors:      CUDA0 buffer size =  4807.06 MiB
...................................................................................................
llama_new_context_with_model: n_ctx      = 8192
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  = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:      CUDA0 KV buffer size =  1024.00 MiB
llama_new_context_with_model: KV self size  = 1024.00 MiB, K (f16):  512.00 MiB, V (f16):  512.00 MiB
llama_new_context_with_model:  CUDA_Host  output buffer size =     0.24 MiB
llama_new_context_with_model:      CUDA0 compute buffer size =    88.00 MiB
llama_new_context_with_model:  CUDA_Host compute buffer size =    24.01 MiB
llama_new_context_with_model: graph nodes  = 903
llama_new_context_with_model: graph splits = 2
[1714720299] warming up the model with an empty run
D:\Arbeit\windows_llama.cpp\vendor\llama.cpp\ggml-cuda\fattn.cu:571: ERROR: CUDA kernel flash_attn_ext_f16 has no device code compatible with CUDA arch 610. ggml-cuda.cu was compiled for: 520,610,700
[...]
D:\Arbeit\windows_llama.cpp\vendor\llama.cpp\ggml-cuda\fattn.cu:571: ERROR: CUDA kernel flash_attn_ext_f16 has no device code compatible with CUDA arch 610. ggml-cuda.cu was compiled for: 520,610,700
ggml_cuda_compute_forward: SILU failed
CUDA error: unspecified launch failure
  current device: 0, in function ggml_cuda_compute_forward at D:\Arbeit\windows_llama.cpp\vendor\llama.cpp\ggml-cuda.cu:2305
  err
GGML_ASSERT: D:\Arbeit\windows_llama.cpp\vendor\llama.cpp\ggml-cuda.cu:61: !"CUDA error"

Flash Attention is working for me on a Turing GPU.

Question: Should llama.cpp support Pascal GPUs?

If not it would be nice to have a feature detection and a speaking error message that indicates this.

[JohannesGaessler]

As of right now there are only CUDA FlashAttention kernels that use tensor cores which in turn means you need Turing/Volta or newer.

[countzero]

@JohannesGaessler Okay, that explains things 😉

So basically the Flash Attention implementation of llama.cpp relies on the NVIDIA GPU to have Tensor Cores.

Suggestion

If I read https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#features-and-technical-specifications correctly we can programmatically determine if a GPU has Tensor Cores.

Could the llama.cpp implementation check the Compute Capability of the GPU and ignore the --flash-attn option with a warning message if it is < 7.0 ?

I am getting a Compute Capability of 6.1 on the NVIDIA Quadro P3200 using the following command:

nvidia-smi --query-gpu=compute_cap --format=csv

[JohannesGaessler]

I'm writing FlashAttention kernels that don't need tensor cores as we speak.

[crashr]

@JohannesGaessler I already read on LocalLLaMA. I am smiling in 5xP40.

[scottmudge]

@JohannesGaessler Is this the PR for that change? #7061

EDIT: Nevermind, I assume so since the branch name is cuda-fa-no-tc-5. I'll test it on my Pascal cards.

[JohannesGaessler]

You're going to get correct results but unless you have a P100 the performance is going to be terrible.

[scottmudge]

Ah, good to know. 2x1080 Ti's, which are roughly equivalent, perhaps marginally faster(?) than P100s in some metrics. It will be an interesting experiment anyway.

[JohannesGaessler]

No, it will literally only perform well on P100s because all other Pascal cards have gimped FP16 performance.

[scottmudge]

Ah that's a shame, well that's good to know.

[scottmudge]

Yeah, went from:

(9.1ms/T = 109.51T/s), Generate:17.10s (18.4ms/T = 54.43T/s), Total:17.45s (53.35T/s)

To:

(11.4ms/T = 87.46T/s), Generate:12.72s (13.7ms/T = 73.20T/s), Total:35.45s (26.26T/s)

With FP16 flash-attention on my non-P100 Pascal cards. So not great. I have a 12GB Maxwell card that I don't think has the same NVIDIA FP16 kneecapping, but for now I'll just leave FA off.

[JohannesGaessler]

General Pascal support: #7188 .
Performance for large batch sizes is still suboptimal due to a lack of a specialized kernel.