Skip to content

Eval bug: With Rocm the gemma-4-26B-A4B-it model causes endless loop #21416

@MiklosPathy

Description

@MiklosPathy

Name and Version

./llama-cli --version
ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
load_backend: loaded ROCm backend from /home/mpathy/llama-b8660-rocm/libggml-hip.so
load_backend: loaded RPC backend from /home/mpathy/llama-b8660-rocm/libggml-rpc.so
load_backend: loaded CPU backend from /home/mpathy/llama-b8660-rocm/libggml-cpu-zen4.so
version: 8660 (d006858)
built with GNU 11.4.0 for Linux x86_64

Operating systems

Linux

GGML backends

Vulkan

Hardware

Strix Halo Ryzen AI Max+ 395, 128G

Models

gemma-4-26B-A4B
https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF
BF16 and Q8 quants tested.

Problem description & steps to reproduce

Just start the model, i used the cli:
./llama-server -m ../models/gemma-4-26B-A4B-it-BF16-00001-of-00002.gguf --mmproj ../models/gemma-4-26B-A4B-it-mmproj-F32.gguf -c 0 -fa on --host 0.0.0.0 --jinja --webui-mcp-proxy
It will produce nothing else just this in endless loop:
<|channel><unused24><unused24><unused24><unused24><unused24><unused24><unused24><unused24><unused24><unused24>

With vulkan it works on the same hardware. Also works with cuda on different machine. This looks a rocm implementation issue. I had to lie in the GGML backends* question, because no rocm otion there.
Other models, eg the gemma-4-31 works fine with rocm on the same machine.

First Bad Commit

No response

Relevant log output

Logs
/llama-b8660-rocm$ ./llama-server -m ../models/gemma-4-26B-A4B-it-BF16-00001-of-00002.gguf --mmproj ../models/gemma-4-26B-A4B-it-mmproj-F32.gguf -c 0 -fa on --host 0.0.0.0 --jinja --webui-mcp-proxy
ggml_cuda_init: found 1 ROCm devices (Total VRAM: 126976 MiB):
  Device 0: AMD Radeon Graphics, gfx1151 (0x1151), VMM: no, Wave Size: 32, VRAM: 126976 MiB
load_backend: loaded ROCm backend from /home/mpathy/llama-b8660-rocm/libggml-hip.so
load_backend: loaded RPC backend from /home/mpathy/llama-b8660-rocm/libggml-rpc.so
load_backend: loaded CPU backend from /home/mpathy/llama-b8660-rocm/libggml-cpu-zen4.so
main: n_parallel is set to auto, using n_parallel = 4 and kv_unified = true
system info: n_threads = 16, n_threads_batch = 16, total_threads = 32

system_info: n_threads = 16 (n_threads_batch = 16) / 32 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |

Running without SSL
init: using 31 threads for HTTP server
srv          main: -----------------
srv          main: CORS proxy is enabled, do not expose server to untrusted environments
srv          main: This feature is EXPERIMENTAL and may be removed or changed in future versions
srv          main: -----------------
start: binding port with default address family
main: loading model
srv    load_model: loading model '../models/gemma-4-26B-A4B-it-BF16-00001-of-00002.gguf'
common_init_result: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on
llama_params_fit_impl: projected to use 54994 MiB of device memory vs. 125151 MiB of free device memory
llama_params_fit_impl: will leave 70156 >= 1024 MiB of free device memory, no changes needed
llama_params_fit: successfully fit params to free device memory
llama_params_fit: fitting params to free memory took 0.46 seconds
llama_model_load_from_file_impl: using device ROCm0 (AMD Radeon Graphics) (0000:c5:00.0) - 125151 MiB free
llama_model_loader: additional 1 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 55 key-value pairs and 658 tensors from ../models/gemma-4-26B-A4B-it-BF16-00001-of-00002.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              = gemma4
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                     general.sampling.top_k i32              = 64
llama_model_loader: - kv   3:                     general.sampling.top_p f32              = 0.950000
llama_model_loader: - kv   4:                      general.sampling.temp f32              = 1.000000
llama_model_loader: - kv   5:                               general.name str              = Gemma-4-26B-A4B-It
llama_model_loader: - kv   6:                           general.finetune str              = it
llama_model_loader: - kv   7:                           general.basename str              = Gemma-4-26B-A4B-It
llama_model_loader: - kv   8:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   9:                         general.size_label str              = 26B-A4B
llama_model_loader: - kv  10:                            general.license str              = apache-2.0
llama_model_loader: - kv  11:                       general.license.link str              = https://ai.google.dev/gemma/docs/gemm...
llama_model_loader: - kv  12:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv  13:                               general.tags arr[str,1]       = ["image-text-to-text"]
llama_model_loader: - kv  14:                         gemma4.block_count u32              = 30
llama_model_loader: - kv  15:                      gemma4.context_length u32              = 262144
llama_model_loader: - kv  16:                    gemma4.embedding_length u32              = 2816
llama_model_loader: - kv  17:                 gemma4.feed_forward_length u32              = 2112
llama_model_loader: - kv  18:                gemma4.attention.head_count u32              = 16
llama_model_loader: - kv  19:             gemma4.attention.head_count_kv arr[i32,30]      = [8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, ...
llama_model_loader: - kv  20:                      gemma4.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  21:                  gemma4.rope.freq_base_swa f32              = 10000.000000
llama_model_loader: - kv  22:    gemma4.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  23:                        gemma4.expert_count u32              = 128
llama_model_loader: - kv  24:                   gemma4.expert_used_count u32              = 8
llama_model_loader: - kv  25:                gemma4.attention.key_length u32              = 512
llama_model_loader: - kv  26:              gemma4.attention.value_length u32              = 512
llama_model_loader: - kv  27:                          general.file_type u32              = 32
llama_model_loader: - kv  28:             gemma4.final_logit_softcapping f32              = 30.000000
llama_model_loader: - kv  29:            gemma4.attention.sliding_window u32              = 1024
llama_model_loader: - kv  30:          gemma4.attention.shared_kv_layers u32              = 0
llama_model_loader: - kv  31:    gemma4.embedding_length_per_layer_input u32              = 0
llama_model_loader: - kv  32:    gemma4.attention.sliding_window_pattern arr[bool,30]     = [true, true, true, true, true, false,...
llama_model_loader: - kv  33:            gemma4.attention.key_length_swa u32              = 256
llama_model_loader: - kv  34:          gemma4.attention.value_length_swa u32              = 256
llama_model_loader: - kv  35:          gemma4.expert_feed_forward_length u32              = 704
llama_model_loader: - kv  36:                gemma4.rope.dimension_count u32              = 512
llama_model_loader: - kv  37:            gemma4.rope.dimension_count_swa u32              = 256
llama_model_loader: - kv  38:               general.quantization_version u32              = 2
llama_model_loader: - kv  39:                       tokenizer.ggml.model str              = gemma4
llama_model_loader: - kv  40:                      tokenizer.ggml.tokens arr[str,262144]  = ["<pad>", "<eos>", "<bos>", "<unk>", ...
llama_model_loader: - kv  41:                      tokenizer.ggml.scores arr[f32,262144]  = [-1000.000000, -1000.000000, -1000.00...
llama_model_loader: - kv  42:                  tokenizer.ggml.token_type arr[i32,262144]  = [3, 1, 3, 3, 3, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  43:                      tokenizer.ggml.merges arr[str,514906]  = ["\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n \n", ...
llama_model_loader: - kv  44:                tokenizer.ggml.bos_token_id u32              = 2
llama_model_loader: - kv  45:                tokenizer.ggml.eos_token_id u32              = 106
llama_model_loader: - kv  46:            tokenizer.ggml.unknown_token_id u32              = 3
llama_model_loader: - kv  47:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  48:               tokenizer.ggml.mask_token_id u32              = 4
llama_model_loader: - kv  49:                    tokenizer.chat_template str              = {%- macro format_parameters(propertie...
llama_model_loader: - kv  50:            tokenizer.ggml.add_space_prefix bool             = false
llama_model_loader: - kv  51:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  52:                                   split.no u16              = 0
llama_model_loader: - kv  53:                                split.count u16              = 2
llama_model_loader: - kv  54:                        split.tensors.count i32              = 658
llama_model_loader: - type  f32:  392 tensors
llama_model_loader: - type bf16:  266 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = BF16
print_info: file size   = 47.02 GiB (16.01 BPW)
load: 0 unused tokens
load: control-looking token:    212 '</s>' was not control-type; this is probably a bug in the model. its type will be overridden
load: printing all EOG tokens:
load:   - 106 ('<turn|>')
load:   - 212 ('</s>')
load: special tokens cache size = 24
load: token to piece cache size = 1.9443 MB
print_info: arch                  = gemma4
print_info: vocab_only            = 0
print_info: no_alloc              = 0
print_info: n_ctx_train           = 262144
print_info: n_embd                = 2816
print_info: n_embd_inp            = 2816
print_info: n_layer               = 30
print_info: n_head                = 16
print_info: n_head_kv             = [8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2, 8, 8, 8, 8, 8, 2]
print_info: n_rot                 = 512
print_info: n_swa                 = 1024
print_info: is_swa_any            = 1
print_info: n_embd_head_k         = 512
print_info: n_embd_head_v         = 512
print_info: n_gqa                 = [2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8, 2, 2, 2, 2, 2, 8]
print_info: n_embd_k_gqa          = [2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024]
print_info: n_embd_v_gqa          = [2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 1024, 2048, 2048, 2048, 2048, 2048, 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: f_attn_scale          = 1.0e+00
print_info: n_ff                  = 2112
print_info: n_expert              = 128
print_info: n_expert_used         = 8
print_info: n_expert_groups       = 0
print_info: n_group_used          = 0
print_info: causal attn           = 1
print_info: pooling type          = -1
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: freq_base_swa         = 10000.0
print_info: freq_scale_swa        = 1
print_info: n_embd_head_k_swa     = 256
print_info: n_embd_head_v_swa     = 256
print_info: n_rot_swa             = 256
print_info: n_ctx_orig_yarn       = 262144
print_info: rope_yarn_log_mul     = 0.0000
print_info: rope_finetuned        = unknown
print_info: model type            = ?B
print_info: model params          = 25.23 B
print_info: general.name          = Gemma-4-26B-A4B-It
print_info: vocab type            = BPE
print_info: n_vocab               = 262144
print_info: n_merges              = 514906
print_info: BOS token             = 2 '<bos>'
print_info: EOS token             = 106 '<turn|>'
print_info: UNK token             = 3 '<unk>'
print_info: PAD token             = 0 '<pad>'
print_info: MASK token            = 4 '<mask>'
print_info: LF token              = 107 '
'
print_info: EOG token             = 106 '<turn|>'
print_info: EOG token             = 212 '</s>'
print_info: max token length      = 93
load_tensors: loading model tensors, this can take a while... (mmap = true, direct_io = false)
str: cannot properly format tensor name output with suffix=weight bid=-1 xid=-1
load_tensors: offloading output layer to GPU
load_tensors: offloading 29 repeating layers to GPU
load_tensors: offloaded 31/31 layers to GPU
load_tensors:   CPU_Mapped model buffer size =  1408.00 MiB
load_tensors:        ROCm0 model buffer size = 48150.36 MiB
.......................................................................
common_init_result: added <turn|> logit bias = -inf
common_init_result: added </s> logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max     = 4
llama_context: n_ctx         = 262144
llama_context: n_ctx_seq     = 262144
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = true
llama_context: freq_base     = 1000000.0
llama_context: freq_scale    = 1
llama_context:  ROCm_Host  output buffer size =     4.00 MiB
llama_kv_cache_iswa: creating non-SWA KV cache, size = 262144 cells
llama_kv_cache:      ROCm0 KV buffer size =  5120.00 MiB
llama_kv_cache: size = 5120.00 MiB (262144 cells,   5 layers,  4/1 seqs), K (f16): 2560.00 MiB, V (f16): 2560.00 MiB
llama_kv_cache: attn_rot_k = 0
llama_kv_cache: attn_rot_v = 0
llama_kv_cache_iswa: creating     SWA KV cache, size = 4608 cells
llama_kv_cache:      ROCm0 KV buffer size =   900.00 MiB
llama_kv_cache: size =  900.00 MiB (  4608 cells,  25 layers,  4/1 seqs), K (f16):  450.00 MiB, V (f16):  450.00 MiB
llama_kv_cache: attn_rot_k = 0
llama_kv_cache: attn_rot_v = 0
sched_reserve: reserving ...
sched_reserve: resolving fused Gated Delta Net support:
sched_reserve: fused Gated Delta Net (autoregressive) enabled
sched_reserve: fused Gated Delta Net (chunked) enabled
sched_reserve:      ROCm0 compute buffer size =   824.52 MiB
sched_reserve:  ROCm_Host compute buffer size =   532.02 MiB
sched_reserve: graph nodes  = 2647
sched_reserve: graph splits = 2
sched_reserve: reserve took 69.27 ms, sched copies = 1
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
clip_model_loader: model name:   Gemma-4-26B-A4B-It
clip_model_loader: description:
clip_model_loader: GGUF version: 3
clip_model_loader: alignment:    32
clip_model_loader: n_tensors:    356
clip_model_loader: n_kv:         28

clip_model_loader: has vision encoder
clip_ctx: CLIP using ROCm0 backend
load_hparams: projector:          gemma4v
load_hparams: n_embd:             1152
load_hparams: n_head:             16
load_hparams: n_ff:               4304
load_hparams: n_layer:            27
load_hparams: ffn_op:             gelu_quick
load_hparams: projection_dim:     2816

--- vision hparams ---
load_hparams: image_size:         224
load_hparams: patch_size:         16
load_hparams: has_llava_proj:     0
load_hparams: minicpmv_version:   0
load_hparams: n_merge:            3
load_hparams: n_wa_pattern: 0
load_hparams: image_min_pixels:   580608
load_hparams: image_max_pixels:   645120

load_hparams: model size:         2185.04 MiB
load_hparams: metadata size:      0.12 MiB
warmup: warmup with image size = 768 x 768
alloc_compute_meta:      ROCm0 compute buffer size =   140.50 MiB
alloc_compute_meta:        CPU compute buffer size =     6.77 MiB
alloc_compute_meta: graph splits = 1, nodes = 1569
warmup: flash attention is enabled
srv    load_model: loaded multimodal model, '../models/gemma-4-26B-A4B-it-mmproj-F32.gguf'
srv    load_model: initializing slots, n_slots = 4
no implementations specified for speculative decoding
slot   load_model: id  0 | task -1 | speculative decoding context not initialized
slot   load_model: id  0 | task -1 | new slot, n_ctx = 262144
no implementations specified for speculative decoding
slot   load_model: id  1 | task -1 | speculative decoding context not initialized
slot   load_model: id  1 | task -1 | new slot, n_ctx = 262144
no implementations specified for speculative decoding
slot   load_model: id  2 | task -1 | speculative decoding context not initialized
slot   load_model: id  2 | task -1 | new slot, n_ctx = 262144
no implementations specified for speculative decoding
slot   load_model: id  3 | task -1 | speculative decoding context not initialized
slot   load_model: id  3 | task -1 | new slot, n_ctx = 262144
srv    load_model: prompt cache is enabled, size limit: 8192 MiB
srv    load_model: use `--cache-ram 0` to disable the prompt cache
srv    load_model: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
srv          init: init: idle slots will be saved to prompt cache and cleared upon starting a new task
init: chat template, example_format: '<bos><|turn>system
<|think|>You are a helpful assistant<turn|>
<|turn>user
Hello<turn|>
<|turn>model
Hi there<turn|>
<|turn>user
How are you?<turn|>
<|turn>model
'
srv          init: init: chat template, thinking = 1
main: model loaded
main: server is listening on http://0.0.0.0:8080
main: starting the main loop...
srv  update_slots: all slots are idle
srv  proxy_reques: proxying GET request to https://www.google.com:443/s2/favicons?domain=0.1&sz=32
srv  log_server_r: done request: GET /cors-proxy 192.168.88.57 404
srv  params_from_: Chat format: peg-gemma4
slot get_availabl: id  3 | task -1 | selected slot by LRU, t_last = -1
srv  get_availabl: updating prompt cache
srv          load:  - looking for better prompt, base f_keep = -1.000, sim = 0.000
srv        update:  - cache state: 0 prompts, 0.000 MiB (limits: 8192.000 MiB, 262144 tokens, 8589934592 est)
srv  get_availabl: prompt cache update took 0.01 ms
slot launch_slot_: id  3 | task -1 | sampler chain: logits -> ?penalties -> ?dry -> ?top-n-sigma -> top-k -> ?typical -> top-p -> min-p -> ?xtc -> temp-ext -> dist
slot launch_slot_: id  3 | task 0 | processing task, is_child = 0
slot update_slots: id  3 | task 0 | new prompt, n_ctx_slot = 262144, n_keep = 0, task.n_tokens = 18
slot update_slots: id  3 | task 0 | n_tokens = 0, memory_seq_rm [0, end)
srv  log_server_r: done request: POST /v1/chat/completions 192.168.88.57 200
slot update_slots: id  3 | task 0 | prompt processing progress, n_tokens = 14, batch.n_tokens = 14, progress = 0.777778
slot update_slots: id  3 | task 0 | n_tokens = 14, memory_seq_rm [14, end)
slot init_sampler: id  3 | task 0 | init sampler, took 0.01 ms, tokens: text = 18, total = 18
slot update_slots: id  3 | task 0 | prompt processing done, n_tokens = 18, batch.n_tokens = 4
srv          stop: cancel task, id_task = 0
slot      release: id  3 | task 0 | stop processing: n_tokens = 95, truncated = 0
srv  update_slots: all slots are idle

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions