Bug: Docker ROCm crashs, only works on metal compiled.

rudiservo commented 2 weeks ago

What happened?

The docker version with ROCm 5.6 exits after graph splits, I tried building and image with ROCm 5.6, 5.7.1, 6.1.2.

These last ones give me an error that is in the logs.

If I compiled and run it on Metal, it works flawlessly.

I have been trying to run it with several version for the past 7 days.

Name and Version

Latest build, always pulled from the last 7 days.

System is Pop_Os 22.04 ROCm 6.1.2 Kernel 6.9.3

What operating system are you seeing the problem on?

Linux

Relevant log output

llamacpp_1  | INFO [                    main] build info | tid="133799363425664" timestamp=1719689759 build=0 commit="unknown"
llamacpp_1  | INFO [                    main] system info | tid="133799363425664" timestamp=1719689759 n_threads=16 n_threads_batch=-1 total_threads=32 system_info="AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | FMA = 1 | NEON = 0 | SVE = 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 | "
llamacpp_1  | llama_model_loader: loaded meta data with 26 key-value pairs and 291 tensors from /models/Mistral-7B-Instruct-v0.3-Q8_0.gguf (version GGUF V3 (latest))
llamacpp_1  | llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llamacpp_1  | llama_model_loader: - kv   0:                       general.architecture str              = llama
llamacpp_1  | llama_model_loader: - kv   1:                               general.name str              = Mistral-7B-Instruct-v0.3
llamacpp_1  | llama_model_loader: - kv   2:                          llama.block_count u32              = 32
llamacpp_1  | llama_model_loader: - kv   3:                       llama.context_length u32              = 32768
llamacpp_1  | llama_model_loader: - kv   4:                     llama.embedding_length u32              = 4096
llamacpp_1  | llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 14336
llamacpp_1  | llama_model_loader: - kv   6:                 llama.attention.head_count u32              = 32
llamacpp_1  | llama_model_loader: - kv   7:              llama.attention.head_count_kv u32              = 8
llamacpp_1  | llama_model_loader: - kv   8:                       llama.rope.freq_base f32              = 1000000.000000
llamacpp_1  | llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
llamacpp_1  | llama_model_loader: - kv  10:                          general.file_type u32              = 7
llamacpp_1  | llama_model_loader: - kv  11:                           llama.vocab_size u32              = 32768
llamacpp_1  | llama_model_loader: - kv  12:                 llama.rope.dimension_count u32              = 128
llamacpp_1  | llama_model_loader: - kv  13:            tokenizer.ggml.add_space_prefix bool             = true
llamacpp_1  | llama_model_loader: - kv  14:                       tokenizer.ggml.model str              = llama
llamacpp_1  | llama_model_loader: - kv  15:                         tokenizer.ggml.pre str              = default
llamacpp_1  | llama_model_loader: - kv  16:                      tokenizer.ggml.tokens arr[str,32768]   = ["<unk>", "<s>", "</s>", "[INST]", "[...
llamacpp_1  | llama_model_loader: - kv  17:                      tokenizer.ggml.scores arr[f32,32768]   = [0.000000, 0.000000, 0.000000, 0.0000...
llamacpp_1  | llama_model_loader: - kv  18:                  tokenizer.ggml.token_type arr[i32,32768]   = [2, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, ...
llamacpp_1  | llama_model_loader: - kv  19:                tokenizer.ggml.bos_token_id u32              = 1
llamacpp_1  | llama_model_loader: - kv  20:                tokenizer.ggml.eos_token_id u32              = 2
llamacpp_1  | llama_model_loader: - kv  21:            tokenizer.ggml.unknown_token_id u32              = 0
llamacpp_1  | llama_model_loader: - kv  22:               tokenizer.ggml.add_bos_token bool             = true
llamacpp_1  | llama_model_loader: - kv  23:               tokenizer.ggml.add_eos_token bool             = false
llamacpp_1  | llama_model_loader: - kv  24:                    tokenizer.chat_template str              = {{ bos_token }}{% for message in mess...
llamacpp_1  | llama_model_loader: - kv  25:               general.quantization_version u32              = 2
llamacpp_1  | llama_model_loader: - type  f32:   65 tensors
llamacpp_1  | llama_model_loader: - type q8_0:  226 tensors
llamacpp_1  | llm_load_vocab: special tokens cache size = 1027
llamacpp_1  | llm_load_vocab: token to piece cache size = 0.1731 MB
llamacpp_1  | llm_load_print_meta: format           = GGUF V3 (latest)
llamacpp_1  | llm_load_print_meta: arch             = llama
llamacpp_1  | llm_load_print_meta: vocab type       = SPM
llamacpp_1  | llm_load_print_meta: n_vocab          = 32768
llamacpp_1  | llm_load_print_meta: n_merges         = 0
llamacpp_1  | llm_load_print_meta: n_ctx_train      = 32768
llamacpp_1  | llm_load_print_meta: n_embd           = 4096
llamacpp_1  | llm_load_print_meta: n_head           = 32
llamacpp_1  | llm_load_print_meta: n_head_kv        = 8
llamacpp_1  | llm_load_print_meta: n_layer          = 32
llamacpp_1  | llm_load_print_meta: n_rot            = 128
llamacpp_1  | llm_load_print_meta: n_embd_head_k    = 128
llamacpp_1  | llm_load_print_meta: n_embd_head_v    = 128
llamacpp_1  | llm_load_print_meta: n_gqa            = 4
llamacpp_1  | llm_load_print_meta: n_embd_k_gqa     = 1024
llamacpp_1  | llm_load_print_meta: n_embd_v_gqa     = 1024
llamacpp_1  | llm_load_print_meta: f_norm_eps       = 0.0e+00
llamacpp_1  | llm_load_print_meta: f_norm_rms_eps   = 1.0e-05
llamacpp_1  | llm_load_print_meta: f_clamp_kqv      = 0.0e+00
llamacpp_1  | llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llamacpp_1  | llm_load_print_meta: f_logit_scale    = 0.0e+00
llamacpp_1  | llm_load_print_meta: n_ff             = 14336
llamacpp_1  | llm_load_print_meta: n_expert         = 0
llamacpp_1  | llm_load_print_meta: n_expert_used    = 0
llamacpp_1  | llm_load_print_meta: causal attn      = 1
llamacpp_1  | llm_load_print_meta: pooling type     = 0
llamacpp_1  | llm_load_print_meta: rope type        = 0
llamacpp_1  | llm_load_print_meta: rope scaling     = linear
llamacpp_1  | llm_load_print_meta: freq_base_train  = 1000000.0
llamacpp_1  | llm_load_print_meta: freq_scale_train = 1
llamacpp_1  | llm_load_print_meta: n_ctx_orig_yarn  = 32768
llamacpp_1  | llm_load_print_meta: rope_finetuned   = unknown
llamacpp_1  | llm_load_print_meta: ssm_d_conv       = 0
llamacpp_1  | llm_load_print_meta: ssm_d_inner      = 0
llamacpp_1  | llm_load_print_meta: ssm_d_state      = 0
llamacpp_1  | llm_load_print_meta: ssm_dt_rank      = 0
llamacpp_1  | llm_load_print_meta: model type       = 7B
llamacpp_1  | llm_load_print_meta: model ftype      = Q8_0
llamacpp_1  | llm_load_print_meta: model params     = 7.25 B
llamacpp_1  | llm_load_print_meta: model size       = 7.17 GiB (8.50 BPW) 
llamacpp_1  | llm_load_print_meta: general.name     = Mistral-7B-Instruct-v0.3
llamacpp_1  | llm_load_print_meta: BOS token        = 1 '<s>'
llamacpp_1  | llm_load_print_meta: EOS token        = 2 '</s>'
llamacpp_1  | llm_load_print_meta: UNK token        = 0 '<unk>'
llamacpp_1  | llm_load_print_meta: LF token         = 781 '<0x0A>'
llamacpp_1  | llm_load_print_meta: max token length = 48
llamacpp_1  | ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
llamacpp_1  | ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
llamacpp_1  | ggml_cuda_init: found 1 ROCm devices:
llamacpp_1  |   Device 0: Radeon RX 7900 XTX, compute capability 11.0, VMM: no
llamacpp_1  | llm_load_tensors: ggml ctx size =    0.27 MiB
llamacpp_1  | llm_load_tensors: offloading 32 repeating layers to GPU
llamacpp_1  | llm_load_tensors: offloading non-repeating layers to GPU
llamacpp_1  | llm_load_tensors: offloaded 33/33 layers to GPU
llamacpp_1  | llm_load_tensors:      ROCm0 buffer size =  7209.02 MiB
llamacpp_1  | llm_load_tensors:        CPU buffer size =   136.00 MiB
llamacpp_1  | ...................................................................................................
llamacpp_1  | llama_new_context_with_model: n_ctx      = 512
llamacpp_1  | llama_new_context_with_model: n_batch    = 512
llamacpp_1  | llama_new_context_with_model: n_ubatch   = 512
llamacpp_1  | llama_new_context_with_model: flash_attn = 0
llamacpp_1  | llama_new_context_with_model: freq_base  = 1000000.0
llamacpp_1  | llama_new_context_with_model: freq_scale = 1
llamacpp_1  | llama_kv_cache_init:      ROCm0 KV buffer size =    64.00 MiB
llamacpp_1  | llama_new_context_with_model: KV self size  =   64.00 MiB, K (f16):   32.00 MiB, V (f16):   32.00 MiB
llamacpp_1  | llama_new_context_with_model:  ROCm_Host  output buffer size =     0.25 MiB
llamacpp_1  | llama_new_context_with_model:      ROCm0 compute buffer size =    81.00 MiB
llamacpp_1  | llama_new_context_with_model:  ROCm_Host compute buffer size =     9.01 MiB
llamacpp_1  | llama_new_context_with_model: graph nodes  = 1030
llamacpp_1  | llama_new_context_with_model: graph splits = 2
llamacpp_1  | ggml_cuda_compute_forward: RMS_NORM failed
llamacpp_1  | CUDA error: invalid device function
llamacpp_1  |   current device: 0, in function ggml_cuda_compute_forward at ggml/src/ggml-cuda.cu:2285
llamacpp_1  |   err
llamacpp_1  | GGML_ASSERT: ggml/src/ggml-cuda.cu:100: !"CUDA error"

Arvamer commented 1 week ago

I had similar error on Archlinux (rocm 6.0.2) and RX 6700 XT and what helped for me is compiling with AMDGPU_TARGETS=gfx1030. Looking at makefile, when AMDGPU_TARGETS is not set, it will auto detect arch as gfx1031. However as gfx1031 is not officially supported. I have to set HSA_OVERRIDE_GFX_VERSION=10.3.0 and I guess it doesn’t like that llama.cpp was compiled for "different" GPU arch.

rudiservo commented 1 week ago

@Arvamer Oh... in the Dockerfile in .devops, the ENV variable that is set is GPU_TARGETS, not AMDGPU_TARGETS.

Going to try and change it, I'll report my findings.

rudiservo commented 1 week ago

Found the issues on the Dockerfile for Rocm. GPU_TARGETS have to be AMDGPU_TARGETS

and ARG ROCM_DOCKER_ARCH is missing " ". So it becomes ARG ROCM_DOCKER_ARCH="\ gfx803 \ gfx900 \ gfx906 \ gfx908 \ gfx90a \ gfx1010 \ gfx1030 \ gfx1100 \ gfx1101 \ gfx1102" There also needs to be 2 different types of rocm versions, Rocm5 and Rocm6. There is a noticeable performance improvement on rocm 6.1.2.

GFX803 and GFX900 are not supported, and GFX906 is deprecated on rocm6.

Should I make a PR?

ggerganov / llama.cpp