Eval bug: GPU Hang Error on Metal backend

[mkbehr]

Name and Version

$ build/bin/llama-cli --version
version: 4857 (0fd7ca7)
built with Apple clang version 16.0.0 (clang-1600.0.26.6) for arm64-apple-darwin24.3.0

Operating systems

Mac

GGML backends

Metal

Hardware

Apple M4 Max

Models

Google Gemma-2 it GGUF

Problem description & steps to reproduce

When I run llama-cli, the inference crashes partway through. Sometimes I see "error: Caused GPU Hang Error (00000003:kIOGPUCommandBufferCallbackErrorHang)", and sometimes "error: Discarded (victim of GPU error/recovery) (00000005:kIOGPUCommandBufferCallbackErrorInnocentVictim)".

I first noticed this after I upgraded my OS to Sequoia 15.3.1. My existing Ollama install started showing these errors. I built llama.cpp from source and replicated them here. So far I've seen the problem on Gemma-2b, Gemma2-2b, and Llama-3.3. I've tried running the Apple hardware diagnostic in case this is a hardware problem, but the diagnostic didn't find any problems.

First Bad Commit

No response

Relevant log output

$ build/bin/llama-cli -m  ../ggml/build/models/gemma-2b-it.gguf -p "This is a test query"
build: 4857 (0fd7ca7a) with Apple clang version 16.0.0 (clang-1600.0.26.6) for arm64-apple-darwin24.3.0
main: llama backend init
main: load the model and apply lora adapter, if any
llama_model_load_from_file_impl: using device Metal (Apple M4 Max) - 98303 MiB free
llama_model_loader: loaded meta data with 19 key-value pairs and 164 tensors from ../ggml/build/models/gemma-2b-it.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              = gemma
llama_model_loader: - kv   1:                               general.name str              = gemma-2b-it
llama_model_loader: - kv   2:                       gemma.context_length u32              = 8192
llama_model_loader: - kv   3:                          gemma.block_count u32              = 18
llama_model_loader: - kv   4:                     gemma.embedding_length u32              = 2048
llama_model_loader: - kv   5:                  gemma.feed_forward_length u32              = 16384
llama_model_loader: - kv   6:                 gemma.attention.head_count u32              = 8
llama_model_loader: - kv   7:              gemma.attention.head_count_kv u32              = 1
llama_model_loader: - kv   8:                 gemma.attention.key_length u32              = 256
llama_model_loader: - kv   9:               gemma.attention.value_length u32              = 256
llama_model_loader: - kv  10:     gemma.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  11:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  12:                tokenizer.ggml.bos_token_id u32              = 2
llama_model_loader: - kv  13:                tokenizer.ggml.eos_token_id u32              = 1
llama_model_loader: - kv  14:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  15:            tokenizer.ggml.unknown_token_id u32              = 3
llama_model_loader: - kv  16:                      tokenizer.ggml.tokens arr[str,256128]  = ["<pad>", "<eos>", "<bos>", "<unk>", ...
llama_model_loader: - kv  17:                      tokenizer.ggml.scores arr[f32,256128]  = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  18:                  tokenizer.ggml.token_type arr[i32,256128]  = [3, 3, 3, 2, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - type  f32:  164 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = all F32 (guessed)
print_info: file size   = 9.34 GiB (32.00 BPW)
load: control-looking token:    107 '<end_of_turn>' was not control-type; this is probably a bug in the model. its type will be overridden
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: special tokens cache size = 5
load: token to piece cache size = 1.6014 MB
print_info: arch             = gemma
print_info: vocab_only       = 0
print_info: n_ctx_train      = 8192
print_info: n_embd           = 2048
print_info: n_layer          = 18
print_info: n_head           = 8
print_info: n_head_kv        = 1
print_info: n_rot            = 256
print_info: n_swa            = 0
print_info: n_embd_head_k    = 256
print_info: n_embd_head_v    = 256
print_info: n_gqa            = 8
print_info: n_embd_k_gqa     = 256
print_info: n_embd_v_gqa     = 256
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             = 16384
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  = 10000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 8192
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       = 2B
print_info: model params     = 2.51 B
print_info: general.name     = gemma-2b-it
print_info: vocab type       = SPM
print_info: n_vocab          = 256128
print_info: n_merges         = 0
print_info: BOS token        = 2 '<bos>'
print_info: EOS token        = 1 '<eos>'
print_info: EOT token        = 107 '<end_of_turn>'
print_info: UNK token        = 3 '<unk>'
print_info: PAD token        = 0 '<pad>'
print_info: LF token         = 227 '<0x0A>'
print_info: EOG token        = 1 '<eos>'
print_info: EOG token        = 107 '<end_of_turn>'
print_info: max token length = 93
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 18 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 19/19 layers to GPU
load_tensors: Metal_Mapped model buffer size =  9561.30 MiB
load_tensors:   CPU_Mapped model buffer size =  2001.00 MiB
.............................................................
llama_init_from_model: n_seq_max     = 1
llama_init_from_model: n_ctx         = 4096
llama_init_from_model: n_ctx_per_seq = 4096
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     = 10000.0
llama_init_from_model: freq_scale    = 1
llama_init_from_model: n_ctx_per_seq (4096) < n_ctx_train (8192) -- the full capacity of the model will not be utilized
ggml_metal_init: allocating
ggml_metal_init: found device: Apple M4 Max
ggml_metal_init: picking default device: Apple M4 Max
ggml_metal_init: using embedded metal library
ggml_metal_init: GPU name:   Apple M4 Max
ggml_metal_init: GPU family: MTLGPUFamilyApple9  (1009)
ggml_metal_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_init: GPU family: MTLGPUFamilyMetal3  (5001)
ggml_metal_init: simdgroup reduction   = true
ggml_metal_init: simdgroup matrix mul. = true
ggml_metal_init: has residency sets    = true
ggml_metal_init: has bfloat            = true
ggml_metal_init: use bfloat            = false
ggml_metal_init: hasUnifiedMemory      = true
ggml_metal_init: recommendedMaxWorkingSetSize  = 103079.22 MB
ggml_metal_init: skipping kernel_get_rows_bf16                     (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_f32                   (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_f32_1row              (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_f32_l4                (not supported)
ggml_metal_init: skipping kernel_mul_mv_bf16_bf16                  (not supported)
ggml_metal_init: skipping kernel_mul_mv_id_bf16_f32                (not supported)
ggml_metal_init: skipping kernel_mul_mm_bf16_f32                   (not supported)
ggml_metal_init: skipping kernel_mul_mm_id_bf16_f32                (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h64           (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h80           (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h96           (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h112          (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h128          (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_bf16_h256          (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_vec_bf16_h128      (not supported)
ggml_metal_init: skipping kernel_flash_attn_ext_vec_bf16_h256      (not supported)
ggml_metal_init: skipping kernel_cpy_f32_bf16                      (not supported)
ggml_metal_init: skipping kernel_cpy_bf16_f32                      (not supported)
ggml_metal_init: skipping kernel_cpy_bf16_bf16                     (not supported)
llama_kv_cache_init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 18, can_shift = 1
llama_kv_cache_init:      Metal KV buffer size =    72.00 MiB
llama_init_from_model: KV self size  =   72.00 MiB, K (f16):   36.00 MiB, V (f16):   36.00 MiB
llama_init_from_model:        CPU  output buffer size =     0.98 MiB
llama_init_from_model:      Metal compute buffer size =   504.25 MiB
llama_init_from_model:        CPU compute buffer size =    12.01 MiB
llama_init_from_model: graph nodes  = 601
llama_init_from_model: graph splits = 2
common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 12

system_info: n_threads = 12 (n_threads_batch = 12) / 16 | Metal : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | AARCH64_REPACK = 1 |

sampler seed: 839441718
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 = 4096
	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 = 4096, n_batch = 2048, n_predict = -1, n_keep = 1

 This is a test query on the Elasticsearch cluster. It is designed to check the health of the cluster and ensure that it is running properly.

ggml_metal_graph_compute: command buffer 0 failed with status 5
error: Caused GPU Hang Error (00000003:kIOGPUCommandBufferCallbackErrorHang)
llama_graph_compute: ggml_backend_sched_graph_compute_async failed with error -1
llama_decode: failed to decode, ret = -3
main : failed to eval
ggml_metal_free: deallocating

[mkbehr]

This turned out to be a hardware issue.