Eval bug: unsloth/Qwen3-Coder-480B-A35B-Instruct-GGUF:Q2_K_XL using HIP backend (AMD MI300X) outputs `GGGGG`

[yeahdongcn]

Name and Version

root@rocm-jupyter-gpu-mi300x1-192gb-devcloud-atl1:~/llama.cpp# ./build/bin/llama-cli --version
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
  Device 0: AMD Instinct MI300X VF, gfx942:sramecc+:xnack- (0x942), VMM: no, Wave Size: 64
version: 5964 (acd6cb1c)
built with cc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0 for x86_64-linux-gnu

Operating systems

Linux

GGML backends

HIP

Hardware

AMD MI300X

Models

unsloth/Qwen3-Coder-480B-A35B-Instruct-GGUF:Q2_K_XL

Problem description & steps to reproduce

./build/bin/llama-cli -hf unsloth/Qwen3-Coder-480B-A35B-Instruct-GGUF:Q2_K_XL -ngl 999

First Bad Commit

No response

Relevant log output

root@rocm-jupyter-gpu-mi300x1-192gb-devcloud-atl1:~/llama.cpp# ./build/bin/llama-cli -hf unsloth/Qwen3-Coder-480B-A35B-Instruct-GGUF:Q2_K_XL -ngl 999
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 ROCm devices:
  Device 0: AMD Instinct MI300X VF, gfx942:sramecc+:xnack- (0x942), VMM: no, Wave Size: 64
curl_perform_with_retry: HEAD https://huggingface.co/unsloth/Qwen3-Coder-480B-A35B-Instruct-GGUF/resolve/main/UD-Q2_K_XL/Qwen3-Coder-480B-A35B-Instruct-UD-Q2_K_XL-00001-of-00004.gguf (attempt 1 of 1)...
common_download_file_single: using cached file: /root/.cache/llama.cpp/unsloth_Qwen3-Coder-480B-A35B-Instruct-GGUF_UD-Q2_K_XL_Qwen3-Coder-480B-A35B-Instruct-UD-Q2_K_XL-00001-of-00004.gguf
curl_perform_with_retry: HEAD https://huggingface.co/unsloth/Qwen3-Coder-480B-A35B-Instruct-GGUF/resolve/main/UD-Q2_K_XL/Qwen3-Coder-480B-A35B-Instruct-UD-Q2_K_XL-00003-of-00004.gguf (attempt 1 of 1)...
curl_perform_with_retry: HEAD https://huggingface.co/unsloth/Qwen3-Coder-480B-A35B-Instruct-GGUF/resolve/main/UD-Q2_K_XL/Qwen3-Coder-480B-A35B-Instruct-UD-Q2_K_XL-00004-of-00004.gguf (attempt 1 of 1)...
curl_perform_with_retry: HEAD https://huggingface.co/unsloth/Qwen3-Coder-480B-A35B-Instruct-GGUF/resolve/main/UD-Q2_K_XL/Qwen3-Coder-480B-A35B-Instruct-UD-Q2_K_XL-00002-of-00004.gguf (attempt 1 of 1)...
common_download_file_single: using cached file: /root/.cache/llama.cpp/unsloth_Qwen3-Coder-480B-A35B-Instruct-GGUF_UD-Q2_K_XL_Qwen3-Coder-480B-A35B-Instruct-UD-Q2_K_XL-00002-of-00004.gguf
common_download_file_single: using cached file: /root/.cache/llama.cpp/unsloth_Qwen3-Coder-480B-A35B-Instruct-GGUF_UD-Q2_K_XL_Qwen3-Coder-480B-A35B-Instruct-UD-Q2_K_XL-00003-of-00004.gguf
common_download_file_single: using cached file: /root/.cache/llama.cpp/unsloth_Qwen3-Coder-480B-A35B-Instruct-GGUF_UD-Q2_K_XL_Qwen3-Coder-480B-A35B-Instruct-UD-Q2_K_XL-00004-of-00004.gguf
build: 5964 (acd6cb1c) with cc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0 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 Instinct MI300X VF) - 195958 MiB free
llama_model_loader: additional 3 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 47 key-value pairs and 747 tensors from /root/.cache/llama.cpp/unsloth_Qwen3-Coder-480B-A35B-Instruct-GGUF_UD-Q2_K_XL_Qwen3-Coder-480B-A35B-Instruct-UD-Q2_K_XL-00001-of-00004.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              = qwen3moe
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen3-Coder-480B-A35B-Instruct
llama_model_loader: - kv   3:                           general.finetune str              = Instruct
llama_model_loader: - kv   4:                           general.basename str              = Qwen3-Coder-480B-A35B-Instruct
llama_model_loader: - kv   5:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   6:                         general.size_label str              = 480B-A35B
llama_model_loader: - kv   7:                            general.license str              = apache-2.0
llama_model_loader: - kv   8:                       general.license.link str              = https://huggingface.co/Qwen/Qwen3-Cod...
llama_model_loader: - kv   9:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv  10:                   general.base_model.count u32              = 1
llama_model_loader: - kv  11:                  general.base_model.0.name str              = Qwen3 Coder 480B A35B Instruct
llama_model_loader: - kv  12:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  13:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen3-Cod...
llama_model_loader: - kv  14:                               general.tags arr[str,2]       = ["unsloth", "text-generation"]
llama_model_loader: - kv  15:                       qwen3moe.block_count u32              = 62
llama_model_loader: - kv  16:                    qwen3moe.context_length u32              = 262144
llama_model_loader: - kv  17:                  qwen3moe.embedding_length u32              = 6144
llama_model_loader: - kv  18:               qwen3moe.feed_forward_length u32              = 8192
llama_model_loader: - kv  19:              qwen3moe.attention.head_count u32              = 96
llama_model_loader: - kv  20:           qwen3moe.attention.head_count_kv u32              = 8
llama_model_loader: - kv  21:                    qwen3moe.rope.freq_base f32              = 10000000.000000
llama_model_loader: - kv  22:  qwen3moe.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  23:                 qwen3moe.expert_used_count u32              = 8
llama_model_loader: - kv  24:              qwen3moe.attention.key_length u32              = 128
llama_model_loader: - kv  25:            qwen3moe.attention.value_length u32              = 128
llama_model_loader: - kv  26:                      qwen3moe.expert_count u32              = 160
llama_model_loader: - kv  27:        qwen3moe.expert_feed_forward_length u32              = 2560
llama_model_loader: - kv  28: qwen3moe.expert_shared_feed_forward_length u32              = 0
llama_model_loader: - kv  29:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  30:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  31:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  32:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  33:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  34:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  35:            tokenizer.ggml.padding_token_id u32              = 151654
llama_model_loader: - kv  36:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  37:                    tokenizer.chat_template str              = {% macro render_item_list(item_list, ...
llama_model_loader: - kv  38:               general.quantization_version u32              = 2
llama_model_loader: - kv  39:                          general.file_type u32              = 10
llama_model_loader: - kv  40:                      quantize.imatrix.file str              = Qwen3-Coder-480B-A35B-Instruct-GGUF/i...
llama_model_loader: - kv  41:                   quantize.imatrix.dataset str              = unsloth_calibration_Qwen3-Coder-480B-...
llama_model_loader: - kv  42:             quantize.imatrix.entries_count u32              = 434
llama_model_loader: - kv  43:              quantize.imatrix.chunks_count u32              = 694
llama_model_loader: - kv  44:                                   split.no u16              = 0
llama_model_loader: - kv  45:                        split.tensors.count i32              = 747
llama_model_loader: - kv  46:                                split.count u16              = 4
llama_model_loader: - type  f32:  311 tensors
llama_model_loader: - type q2_K:  124 tensors
llama_model_loader: - type q3_K:   52 tensors
llama_model_loader: - type q4_K:  234 tensors
llama_model_loader: - type q5_K:   15 tensors
llama_model_loader: - type q6_K:   11 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q2_K - Medium
print_info: file size   = 167.91 GiB (3.00 BPW)
load: special tokens cache size = 26
load: token to piece cache size = 0.9311 MB
print_info: arch             = qwen3moe
print_info: vocab_only       = 0
print_info: n_ctx_train      = 262144
print_info: n_embd           = 6144
print_info: n_layer          = 62
print_info: n_head           = 96
print_info: n_head_kv        = 8
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: is_swa_any       = 0
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 12
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: f_attn_scale     = 0.0e+00
print_info: n_ff             = 8192
print_info: n_expert         = 160
print_info: n_expert_used    = 8
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  = 10000000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 262144
print_info: rope_finetuned   = unknown
print_info: model type       = ?B
print_info: model params     = 480.15 B
print_info: general.name     = Qwen3-Coder-480B-A35B-Instruct
print_info: n_ff_exp         = 2560
print_info: vocab type       = BPE
print_info: n_vocab          = 151936
print_info: n_merges         = 151387
print_info: BOS token        = 11 ','
print_info: EOS token        = 151645 '<|im_end|>'
print_info: EOT token        = 151645 '<|im_end|>'
print_info: PAD token        = 151654 '<|vision_pad|>'
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 62 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 63/63 layers to GPU
load_tensors:        ROCm0 model buffer size = 171435.90 MiB
load_tensors:   CPU_Mapped model buffer size =   500.77 MiB
....................................................................................................
llama_context: constructing llama_context
llama_context: non-unified KV cache requires ggml_set_rows() - forcing unified KV cache
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 4096
llama_context: n_ctx_per_seq = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = 0
llama_context: kv_unified    = true
llama_context: freq_base     = 10000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
llama_context:  ROCm_Host  output buffer size =     0.58 MiB
llama_kv_cache_unified:      ROCm0 KV buffer size =   992.00 MiB
llama_kv_cache_unified: size =  992.00 MiB (  4096 cells,  62 layers,  1/ 1 seqs), K (f16):  496.00 MiB, V (f16):  496.00 MiB
llama_kv_cache_unified: LLAMA_SET_ROWS=0, using old ggml_cpy() method for backwards compatibility
llama_context:      ROCm0 compute buffer size =   848.00 MiB
llama_context:  ROCm_Host compute buffer size =    20.01 MiB
llama_context: graph nodes  = 4222
llama_context: graph splits = 2
common_init_from_params: added <|endoftext|> logit bias = -inf
common_init_from_params: added <|im_end|> logit bias = -inf
common_init_from_params: added <|fim_pad|> logit bias = -inf
common_init_from_params: added <|repo_name|> logit bias = -inf
common_init_from_params: added <|file_sep|> logit bias = -inf
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 = 20
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
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 = 20 (n_threads_batch = 20) / 20 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 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 |

main: interactive mode on.
sampler seed: 1072039959
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-n-sigma -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
generate: n_ctx = 4096, 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


> Hi
GGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGG
>

[yeahdongcn]

@danielhanchen I'm guessing this runs fine on the CUDA backend?

[yeahdongcn]

unsloth/Qwen3-Coder-480B-A35B-Instruct-GGUF:Q4_K_XL works as expected.

./build/bin/llama-cli -m /root/.cache/llama.cpp/unsloth_Qwen3-Coder-480B-A35B-Instruct-GGUF_UD-Q4_K_XL_Qwen3-Coder-480B-A35B-Instruct-UD-Q4_K_XL-00001-of-00006.gguf -ngl 40

[danielhanchen]

@yeahdongcn Yes so I tried Q2_K_XL via ./llama.cpp/llama-cli --model Qwen3-Coder-480B-A35B-Instruct-GGUF/UD-Q2_K_XL/Qwen3-Coder-480B-A35B-Instruct-UD-Q2_K_XL-00001-of-00004.gguf -ngl 99 with "Hi" as what you tried on CUDA:

> Hi
Hello! How can I help you today?

> 1+1
1 + 1 = 2

> 100 / 100                
100 / 100 = 1

> Mini function for Fib in Python
Here's a mini Fibonacci function in Python:

``python
def fib(n):
    if n <= 1:
        return n
    a, b = 0, 1
    for _ in range(2, n + 1):
        a, b = b, a + b
    return b

# Example usage:
print(fib(10))  # Output: 55
``

This is an efficient iterative version that calculates the nth Fibonacci number in O(n) time and O(1) space.

I also tried a new IQ1_M I just made (160GB) via ./llama.cpp/llama-cli --model Qwen3-Coder-480B-A35B-Instruct-GGUF/UD-IQ1_M/Qwen3-Coder-480B-A35B-Instruct-UD-IQ1_M-00001-of-00004.gguf -ngl 99 and I get:

> Hi
Hello! How can I help you today?

> What is the sqrt of 101
The square root of 101 is approximately **10.05**.

To be more precise:
√101 ≈ 10.04987562112089

Since 101 is not a perfect square, its square root is an irrational number that goes on infinitely without repeating.

[ggerganov]

My guess is that this is another instance of #13607. The trend is that new models will require higher FP range and F16 accumulators will not be enough.