Eval bug: inference of 32B eats too much memory on ROCM HIP (5x AMD Radeon Instinct Mi50 (gfx906))

[bene-ges]

Name and Version

./llama-cli --version
ROCm calling rocblas_initialize as a workaround for a rocBLAS bug
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 5 ROCm devices:
  Device 0: AMD Radeon VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 1: AMD Radeon VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 2: AMD Radeon VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 3: AMD Radeon VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 4: AMD Radeon VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
version: 1 (10f2e81)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu

Operating systems

Linux

GGML backends

HIP

Hardware

8 * Intel(R) Core(TM) i7-7700 CPU @ 3.60GHz

5 * AMD Radeon Instinct Mi50 (gfx906)

Models

Qwen2.5-32B-Instruct-Q4_K_M.gguf

Problem description & steps to reproduce

when I run ./llama-cli with context -c 32768 and big prompt, it eats too much VRAM and fails with OOM. The same run on Nvidia gpus, works fine.

This is my exact command.

./llama-cli -m models/Qwen2.5-32B-Instruct-Q4_K_M.gguf --n-gpu-layers 100 -c 32768 -f full_prompt.txt

What happens: first the model is loaded it occupies 50% VRAM on all 5 gpus.
Then it starts reading the prompt, VRAM starts to gradually grow, eating all VRAM and ends with error.

llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:73: ROCm errorROCm error: out of memory

Same behaviour is observed when using llama-server.

Build command

HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)"     cmake -S . -B build -DGGML_CUDA_FA_ALL_QUANTS=ON -DLLAMA_CURL=ON -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx906 -DCMAKE_BUILD_TYPE=Release \

I also add LLAMA_CUDA_NO_PEER_COPY=1 to prevent gibberish output, as mentioned in #3051, but it has no effect on VRAM eating problem.

I tried to run the same inference on nvidia gpus, it worked fine, at the stage of prompt reading VRAM only grows by 10-15% compared to model loading stage, and then stops growing.

First Bad Commit

No response

Relevant log output

ROCm calling rocblas_initialize as a workaround for a rocBLAS bug
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 5 ROCm devices:
  Device 0: AMD Radeon VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 1: AMD Radeon VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 2: AMD Radeon VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 3: AMD Radeon VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
  Device 4: AMD Radeon VII, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64
build: 1 (10f2e81) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device ROCm0 (AMD Radeon VII) - 15922 MiB free
llama_model_load_from_file_impl: using device ROCm1 (AMD Radeon VII) - 16348 MiB free
llama_model_load_from_file_impl: using device ROCm2 (AMD Radeon VII) - 16348 MiB free
llama_model_load_from_file_impl: using device ROCm3 (AMD Radeon VII) - 16348 MiB free
llama_model_load_from_file_impl: using device ROCm4 (AMD Radeon VII) - 16348 MiB free
llama_model_loader: loaded meta data with 38 key-value pairs and 771 tensors from models/Qwen2.5-32B-Instruct-Q4_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              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen2.5 32B Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Qwen2.5
llama_model_loader: - kv   5:                         general.size_label str              = 32B
llama_model_loader: - kv   6:                            general.license str              = apache-2.0
llama_model_loader: - kv   7:                       general.license.link str              = https://huggingface.co/Qwen/Qwen2.5-3...
llama_model_loader: - kv   8:                   general.base_model.count u32              = 1
llama_model_loader: - kv   9:                  general.base_model.0.name str              = Qwen2.5 32B
llama_model_loader: - kv  10:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  11:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen2.5-32B
llama_model_loader: - kv  12:                               general.tags arr[str,2]       = ["chat", "text-generation"]
llama_model_loader: - kv  13:                          general.languages arr[str,1]       = ["en"]
llama_model_loader: - kv  14:                          qwen2.block_count u32              = 64
llama_model_loader: - kv  15:                       qwen2.context_length u32              = 32768
llama_model_loader: - kv  16:                     qwen2.embedding_length u32              = 5120
llama_model_loader: - kv  17:                  qwen2.feed_forward_length u32              = 27648
llama_model_loader: - kv  18:                 qwen2.attention.head_count u32              = 40
llama_model_loader: - kv  19:              qwen2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  20:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  21:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  22:                          general.file_type u32              = 15
llama_model_loader: - kv  23:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  24:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  25:                      tokenizer.ggml.tokens arr[str,152064]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  26:                  tokenizer.ggml.token_type arr[i32,152064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  27:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  28:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  29:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  30:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  31:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  32:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  33:               general.quantization_version u32              = 2
llama_model_loader: - kv  34:                      quantize.imatrix.file str              = /models_out/Qwen2.5-32B-Instruct-GGUF...
llama_model_loader: - kv  35:                   quantize.imatrix.dataset str              = /training_dir/calibration_datav3.txt
llama_model_loader: - kv  36:             quantize.imatrix.entries_count i32              = 448
llama_model_loader: - kv  37:              quantize.imatrix.chunks_count i32              = 128
llama_model_loader: - type  f32:  321 tensors
llama_model_loader: - type q4_K:  385 tensors
llama_model_loader: - type q6_K:   65 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 18.48 GiB (4.85 BPW)
load: special tokens cache size = 22
load: token to piece cache size = 0.9310 MB
print_info: arch             = qwen2
print_info: vocab_only       = 0
print_info: n_ctx_train      = 32768
print_info: n_embd           = 5120
print_info: n_layer          = 64
print_info: n_head           = 40
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 5
print_info: n_embd_k_gqa     = 1024
print_info: n_embd_v_gqa     = 1024
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: n_ff             = 27648
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 = 1
print_info: n_ctx_orig_yarn  = 32768
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       = 32B
print_info: model params     = 32.76 B
print_info: general.name     = Qwen2.5 32B Instruct
print_info: vocab type       = BPE
print_info: n_vocab          = 152064
print_info: n_merges         = 151387
print_info: BOS token        = 151643 '<|endoftext|>'
print_info: EOS token        = 151645 '<|im_end|>'
print_info: EOT token        = 151645 '<|im_end|>'
print_info: PAD token        = 151643 '<|endoftext|>'
print_info: LF token         = 198 'Ċ'
print_info: FIM PRE token    = 151659 '<|fim_prefix|>'
print_info: FIM SUF token    = 151661 '<|fim_suffix|>'
print_info: FIM MID token    = 151660 '<|fim_middle|>'
print_info: FIM PAD token    = 151662 '<|fim_pad|>'
print_info: FIM REP token    = 151663 '<|repo_name|>'
print_info: FIM SEP token    = 151664 '<|file_sep|>'
print_info: EOG token        = 151643 '<|endoftext|>'
print_info: EOG token        = 151645 '<|im_end|>'
print_info: EOG token        = 151662 '<|fim_pad|>'
print_info: EOG token        = 151663 '<|repo_name|>'
print_info: EOG token        = 151664 '<|file_sep|>'
print_info: max token length = 256
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 64 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 65/65 layers to GPU
load_tensors:        ROCm0 model buffer size =  3726.02 MiB
load_tensors:        ROCm1 model buffer size =  3581.64 MiB
load_tensors:        ROCm2 model buffer size =  3545.55 MiB
load_tensors:        ROCm3 model buffer size =  3545.55 MiB
load_tensors:        ROCm4 model buffer size =  4109.59 MiB
load_tensors:   CPU_Mapped model buffer size =   417.66 MiB
................................................................................................
llama_init_from_model: n_seq_max     = 1
llama_init_from_model: n_ctx         = 32768
llama_init_from_model: n_ctx_per_seq = 32768
llama_init_from_model: n_batch       = 2048
llama_init_from_model: n_ubatch      = 512
llama_init_from_model: flash_attn    = 0
llama_init_from_model: freq_base     = 1000000.0
llama_init_from_model: freq_scale    = 1
llama_kv_cache_init: kv_size = 32768, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 64, can_shift = 1
llama_kv_cache_init:      ROCm0 KV buffer size =  1664.00 MiB
llama_kv_cache_init:      ROCm1 KV buffer size =  1664.00 MiB
llama_kv_cache_init:      ROCm2 KV buffer size =  1664.00 MiB
llama_kv_cache_init:      ROCm3 KV buffer size =  1664.00 MiB
llama_kv_cache_init:      ROCm4 KV buffer size =  1536.00 MiB
llama_init_from_model: KV self size  = 8192.00 MiB, K (f16): 4096.00 MiB, V (f16): 4096.00 MiB
llama_init_from_model:  ROCm_Host  output buffer size =     0.58 MiB
llama_init_from_model:      ROCm0 compute buffer size =  2664.00 MiB
llama_init_from_model:      ROCm1 compute buffer size =  2664.00 MiB
llama_init_from_model:      ROCm2 compute buffer size =  2664.00 MiB
llama_init_from_model:      ROCm3 compute buffer size =  2664.00 MiB
llama_init_from_model:      ROCm4 compute buffer size =  2664.00 MiB
llama_init_from_model:  ROCm_Host compute buffer size =    74.01 MiB
llama_init_from_model: graph nodes  = 2246
llama_init_from_model: graph splits = 6
common_init_from_params: setting dry_penalty_last_n to ctx_size = 32768
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 4
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
*** User-specified prompt will pre-start conversation, did you mean to set --system-prompt (-sys) instead?
main: chat template example:
<|im_start|>system
You are a helpful assistant<|im_end|>
<|im_start|>user
Hello<|im_end|>
<|im_start|>assistant
Hi there<|im_end|>
<|im_start|>user
How are you?<|im_end|>
<|im_start|>assistant


system_info: n_threads = 4 (n_threads_batch = 4) / 8 | ROCm : NO_VMM = 1 | NO_PEER_COPY = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 |

main: interactive mode on.
sampler seed: 1112466201
sampler params:
        repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
        dry_multiplier = 0.000, dry_base = 1.750, dry_allowed_length = 2, dry_penalty_last_n = 32768
        top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, top_n_sigma = -1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
generate: n_ctx = 32768, n_batch = 2048, n_predict = -1, n_keep = 0

== Running in interactive mode. ==
 - Press Ctrl+C to interject at any time.
 - Press Return to return control to the AI.
 - To return control without starting a new line, end your input with '/'.
 - If you want to submit another line, end your input with '\'.
 - Not using system message. To change it, set a different value via -sys PROMPT

/home/rig/llamafix/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:73: ROCm error
ROCm error: out of memory
  current device: 4, in function alloc at /home/rig/llamafix/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:366
  ggml_cuda_device_malloc(&ptr, look_ahead_size, device)
Could not attach to process.  If your uid matches the uid of the target
process, check the setting of /proc/sys/kernel/yama/ptrace_scope, or try
again as the root user.  For more details, see /etc/sysctl.d/10-ptrace.conf
ptrace: Inappropriate ioctl for device.
No stack.
The program is not being run.
./run_llama_cli.sh: line 1: 24869 Aborted                 (core dumped) ./llama-cli -m models/Qwen2.5-32B-Instruct-Q4_K_M.gguf --n-gpu-layers 100 -c 32768 -f full_prompt.txt

[bene-ges]

Tried to build llama.cpp with Vulcan, on the same 5x AMD Radeon Instinct Mi50 the same inference command works fine but 2x slower, the VRAM doesn't grow at prompt reading.

[IMbackK]

I can not repoduce this with a trio of gfx908 devices as well as gfx1030, rocm 6.3.2, latest master lamacpp and ./llama-cli -m models/Qwen2.5-32B-Instruct-Q4_K_M.gguf --n-gpu-layers 100 -c 32768

There is no difference in the amount allocated regardless of if there is no prompt or if i start with a 16K long prompt via -f. Its possible that the issue is in your rocm environment or in a code path not hit by gfx908 or 1030. Could you get a hsa trace with rocprof? that would show the allocations.

[bene-ges]

i use rocm-5.7.3 and llama.cpp head commit is 10f2e81

i launched rocprof (see attachment)

rocprof.zip

[IMbackK]

I will take a look, but 5.7.3 is very old and I am aware of several issues with this version, I would strongly suggest upgrading to at least 6.2

[bene-ges]

hi @IMbackK,
i tested with rocm 6.3, got the same behaviour, eats all VRAM and fails with OOM

rocprof files attached

rocprof_6.3.zip

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.