GPU offloading not working on system with AMD 5900HX CPU

[vlasky]

I'm running llamafile 0.8.1 on a Windows 10 mini PC with a AMD Ryzen 9 5900HX CPU

CPU Architecture: AMD Cezanne (Zen 3, Ryzen 5000)
GPU: AMD Radeon RX Vega 8

The mini PC has 64GB RAM installed.

When I enable llamafile GPU support with -ngl 9999, it exits with the error

ggml_cuda_compute_forward: RMS_NORM failed
CUDA error: invalid device function
  current device: 0, in function ggml_cuda_compute_forward at ggml-cuda.cu:11444
  err
GGML_ASSERT: ggml-cuda.cu:9198: !"CUDA error"

My command line is:

llamafile-0.8.1.exe -ngl 9999 -m dolphin-2.9-llama3-8b-Q5_K_M.gguf

I have also tried re-running after installing the AMD HIP SDK but this made no difference.
Contrary to the runtime messages, amdclang++.exe was in my Windows PATH.

import_cuda_impl: initializing gpu module...
get_rocm_bin_path: note: amdclang++.exe not found on $PATH
link_cuda_dso: note: dynamically linking /C/Users/Vlad/.llamafile/ggml-rocm.dll
ggml_cuda_link: welcome to ROCm SDK with tinyBLAS
link_cuda_dso: GPU support loaded
{"build":1500,"commit":"a30b324","function":"server_cli","level":"INFO","line":2858,"msg":"build info","tid":"9442720","timestamp":1714373224}
{"function":"server_cli","level":"INFO","line":2861,"msg":"system info","n_threads":8,"n_threads_batch":-1,"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 = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LAMMAFILE = 1 | ","tid":"9442720","timestamp":1714373224,"total_threads":16}
llama_model_loader: loaded meta data with 22 key-value pairs and 291 tensors from dolphin-2.9-llama3-8b-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              = dolphin-2.9-llama3-8b
llama_model_loader: - kv   2:                          llama.block_count u32              = 32
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.feed_forward_length u32              = 14336
llama_model_loader: - kv   6:                 llama.attention.head_count u32              = 32
llama_model_loader: - kv   7:              llama.attention.head_count_kv u32              = 8
llama_model_loader: - kv   8:                       llama.rope.freq_base f32              = 500000.000000
llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                          general.file_type u32              = 17
llama_model_loader: - kv  11:                           llama.vocab_size u32              = 128258
llama_model_loader: - kv  12:                 llama.rope.dimension_count u32              = 128
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  14:                      tokenizer.ggml.tokens arr[str,128258]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr[i32,128258]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  16:                      tokenizer.ggml.merges arr[str,280147]  = ["─á ─á", "─á ─á─á─á", "─á─á ─á─á", "...
llama_model_loader: - kv  17:                tokenizer.ggml.bos_token_id u32              = 128000
llama_model_loader: - kv  18:                tokenizer.ggml.eos_token_id u32              = 128256
llama_model_loader: - kv  19:            tokenizer.ggml.padding_token_id u32              = 128001
llama_model_loader: - kv  20:                    tokenizer.chat_template str              = {% if not add_generation_prompt is de...
llama_model_loader: - kv  21:               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 ( 258/128258 ).
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = llama
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 128258
llm_load_print_meta: n_merges         = 280147
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  = 500000.0
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     = 8.03 B
llm_load_print_meta: model size       = 5.33 GiB (5.70 BPW)
llm_load_print_meta: general.name     = dolphin-2.9-llama3-8b
llm_load_print_meta: BOS token        = 128000 '<|begin_of_text|>'
llm_load_print_meta: EOS token        = 128256 '<|im_end|>'
llm_load_print_meta: PAD token        = 128001 '<|end_of_text|>'
llm_load_print_meta: LF token         = 128 'Ä'
llm_load_print_meta: EOT token        = 128256 '<|im_end|>'
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
ggml_cuda_init: found 1 ROCm devices:
  Device 0: AMD Radeon(TM) Graphics, compute capability 9.0, VMM: no
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:      ROCm0 buffer size =  5115.50 MiB
llm_load_tensors:        CPU buffer size =   344.44 MiB
........................................................................................
llama_new_context_with_model: n_ctx      = 512
llama_new_context_with_model: n_batch    = 512
llama_new_context_with_model: n_ubatch   = 512
llama_new_context_with_model: freq_base  = 500000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init:      ROCm0 KV buffer size =    64.00 MiB
llama_new_context_with_model: KV self size  =   64.00 MiB, K (f16):   32.00 MiB, V (f16):   32.00 MiB
llama_new_context_with_model:  ROCm_Host  output buffer size =     0.50 MiB
llama_new_context_with_model:      ROCm0 compute buffer size =   258.50 MiB
llama_new_context_with_model:  ROCm_Host compute buffer size =     9.01 MiB
llama_new_context_with_model: graph nodes  = 1030
llama_new_context_with_model: graph splits = 2
ggml_cuda_compute_forward: RMS_NORM failed
CUDA error: invalid device function
  current device: 0, in function ggml_cuda_compute_forward at ggml-cuda.cu:11444
  err
GGML_ASSERT: ggml-cuda.cu:9198: !"CUDA error"

[sleppyrobot]

Hey I believe integrated GPU are not supported, probably better to run on CPU at this time, by passing -ngl 0 instead of 9999

Also I have seen a few open issues with the same error/warning at the start when using AMD, so am not sure if I should open a new issues.
The line that says : get_rocm_bin_path: note: amdclang++.exe not found on $PATH
The actually file located there is named clang++.exe , in Windows, however on Linux is called amdclang++.exe

Perhaps there could be an Operating System check before looking for amdclang++ or clang++.
or maybe it is something else

[vlasky]

Hey I believe integrated GPU are not supported, probably better to run on CPU at this time, by passing -ngl 0 instead of 9999

OK. I was curious to know whether additional acceleration could be obtained by combining the iGPU with the CPU.

In any case, I reckon the docs should explicitly state that AMD iGPUs are not supported (if they're not). Ideally, llamafile should also report this at runtime.

Also I have seen a few open issues with the same error/warning at the start when using AMD, so am not sure if I should open a new issues. The line that says : get_rocm_bin_path: note: amdclang++.exe not found on $PATH The actually file located there is named clang++.exe , in Windows, however on Linux is called amdclang++.exe

Perhaps there could be an Operating System check before looking for amdclang++ or clang++. or maybe it is something else

Yes. I copied clang++.exe to amdclang++.exe to overcome this. Both executables were in the $PATH, but the get_rocm_bin_path: note: amdclang++.exe not found on $PATH message still appeared.

[Laky2k8]

This bug still persists in 0.8.4 which makes AMD GPU offloading impossible on Windows right now :(
Somewhat of a fix is to revert to 0.7.4, where while it still does not load HIPBlas, it does load TinyBLAS which is much faster than CPU.

[lovenemesis]

You might want to refer to my experience on #188 .

In short, iGPU could work with 0.8.4 as long as:

• Parsing the proper HAS override environment variables. In your Vega based iGPU, it should be something like HSA_OVERRIDE_GFX_VERSION=9.0.6
• Make sure it has sufficient VRAM allocated from the unified memory pool.
• Consider utilizing the tinyBLAS instead of hipBLAS by parsing --nocompile.

[Laky2k8]

On 0.8.4 tinyBLAS works again on my RX6600 GPU which is great because I can run AI models with proper speed again! However the bug that llamafile is searching for amdclang++.exe still persists :(

[Djip007]

Make sure it has sufficient VRAM allocated from the unified memory pool.

At least with the use of LLAMA_HIP_UMA=1 on llama.cpp on linux (don't know for Windows) it is not needed. and with this PR (ggerganov/llama.cpp#7414) it is look to be as fast a with VRAM reserve. (can't config high VRAM so can't test on large LLM...)

But we need to add the possibility to use this config with llamafile.

UPDATE(13/08/2024) : with linux kernel 6.10+ look kernel change allow the use of GGT dynamic RAM and not only VRAM on some AMD APU. It is the case for my Ryzen 7940HS.)

[abhinavsingh]

IIUC, GPU are not getting used on AMD * systems. Wondering if there is any workaround this?