/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
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