Skip to content

Eval bug: MiniMax M2.5 output quality regression and missing <think> token on recent builds #21610

@MisticRain

Description

@MisticRain

Name and Version

b8703-5c4aae66e, prebuilt Vulkan x64 binary for Windows

Operating systems

Windows

GGML backends

Vulkan

Hardware

AMD Ryzen™ AI Max+ 395, Radeon 8060s

Models

MiniMax-M2.5-UD-Q3_K_XL.gguf

Problem description:
MiniMax M2.5 output quality has noticeably degraded on b8703 compared to earlier builds (pre-April 2026). The opening think token is consistently missing.
Tested and ruled out:

LLAMA_ATTN_ROT_DISABLE=1 → attn_rot goes to 0 in logs, issue persists
Removing KV cache quantization (f16 instead of q8_0) → attn_rot goes to 0, issue persists
GDN log lines ("fused Gated Delta Net enabled") appear on both working and broken builds → not the cause?
Last known working build was pre-attn_rot merge and it was build llama-b8338-bin-win-vulkan-x64

First Bad Commit

No response

Relevant log output

Logs
.\llama-server.exe  --host 0.0.0.0 --port 8080 -m MiniMax-M2.5-UD-Q3_K_XL-00001-of-00004.gguf -ngl 99 -fa on --cache-type-k q8_0 --cache-type-v q8_0 --top-p 0.95 -t 1.0 --min_p 0.01 -np 1 --top_k 40 -ub 256 --batch-size 512 --jinja --no-mmap
load_backend: loaded RPC backend from C:\Users\AI Max\Downloads\llama-b8703-bin-win-vulkan-x64\ggml-rpc.dll
load_backend: loaded Vulkan backend from C:\Users\AI Max\Downloads\llama-b8703-bin-win-vulkan-x64\ggml-vulkan.dll
load_backend: loaded CPU backend from C:\Users\AI Max\Downloads\llama-b8703-bin-win-vulkan-x64\ggml-cpu-zen4.dll
build_info: b8703-5c4aae66e
system_info: n_threads = 1 (n_threads_batch = 1) / 32 | 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
start: binding port with default address family
main: loading model
srv    load_model: loading model 'MiniMax-M2.5-UD-Q3_K_XL-00001-of-00004.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 121861 MiB of device memory vs. 108781 MiB of free device memory
llama_params_fit_impl: cannot meet free memory target of 1024 MiB, need to reduce device memory by 14104 MiB
llama_params_fit_impl: context size reduced from 196608 to 87296 -> need 14121 MiB less memory in total
llama_params_fit_impl: entire model can be fit by reducing context
llama_params_fit: successfully fit params to free device memory
llama_params_fit: fitting params to free memory took 0.28 seconds
llama_model_load_from_file_impl: using device Vulkan0 (AMD Radeon(TM) 8060S Graphics) (unknown id) - 108782 MiB free
llama_model_loader: additional 3 GGUFs metadata loaded.
llama_model_loader: loaded meta data with 53 key-value pairs and 809 tensors from MiniMax-M2.5-UD-Q3_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              = minimax-m2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                     general.sampling.top_k i32              = 40
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              = Minimax-M2.5
llama_model_loader: - kv   6:                           general.basename str              = Minimax-M2.5
llama_model_loader: - kv   7:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   8:                         general.size_label str              = 256x4.9B
llama_model_loader: - kv   9:                            general.license str              = other
llama_model_loader: - kv  10:                       general.license.name str              = modified-mit
llama_model_loader: - kv  11:                       general.license.link str              = https://github.com/MiniMax-AI/MiniMax...
llama_model_loader: - kv  12:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv  13:                   general.base_model.count u32              = 1
llama_model_loader: - kv  14:                  general.base_model.0.name str              = MiniMax M2.5
llama_model_loader: - kv  15:          general.base_model.0.organization str              = MiniMaxAI
llama_model_loader: - kv  16:              general.base_model.0.repo_url str              = https://huggingface.co/MiniMaxAI/Mini...
llama_model_loader: - kv  17:                               general.tags arr[str,2]       = ["unsloth", "text-generation"]
llama_model_loader: - kv  18:                     minimax-m2.block_count u32              = 62
llama_model_loader: - kv  19:                  minimax-m2.context_length u32              = 196608
llama_model_loader: - kv  20:                minimax-m2.embedding_length u32              = 3072
llama_model_loader: - kv  21:             minimax-m2.feed_forward_length u32              = 1536
llama_model_loader: - kv  22:            minimax-m2.attention.head_count u32              = 48
llama_model_loader: - kv  23:         minimax-m2.attention.head_count_kv u32              = 8
llama_model_loader: - kv  24:                  minimax-m2.rope.freq_base f32              = 5000000.000000
llama_model_loader: - kv  25: minimax-m2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  26:                    minimax-m2.expert_count u32              = 256
llama_model_loader: - kv  27:               minimax-m2.expert_used_count u32              = 8
llama_model_loader: - kv  28:              minimax-m2.expert_gating_func u32              = 2
llama_model_loader: - kv  29:            minimax-m2.attention.key_length u32              = 128
llama_model_loader: - kv  30:          minimax-m2.attention.value_length u32              = 128
llama_model_loader: - kv  31:      minimax-m2.expert_feed_forward_length u32              = 1536
llama_model_loader: - kv  32:            minimax-m2.rope.dimension_count u32              = 64
llama_model_loader: - kv  33:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  34:                         tokenizer.ggml.pre str              = minimax-m2
llama_model_loader: - kv  35:                      tokenizer.ggml.tokens arr[str,200064]  = ["Ā", "ā", "Ă", "ă", "Ą", "ą", ...
llama_model_loader: - kv  36:                  tokenizer.ggml.token_type arr[i32,200064]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  37:                      tokenizer.ggml.merges arr[str,199744]  = ["Ġ Ġ", "Ġ t", "Ġ a", "i n", "e r...
llama_model_loader: - kv  38:                tokenizer.ggml.bos_token_id u32              = 200034
llama_model_loader: - kv  39:                tokenizer.ggml.eos_token_id u32              = 200020
llama_model_loader: - kv  40:            tokenizer.ggml.unknown_token_id u32              = 200021
llama_model_loader: - kv  41:            tokenizer.ggml.padding_token_id u32              = 200004
llama_model_loader: - kv  42:               tokenizer.ggml.add_bos_token bool             = true
llama_model_loader: - kv  43:                    tokenizer.chat_template str              = {# Unsloth template fixes #}\n{# -----...
llama_model_loader: - kv  44:               general.quantization_version u32              = 2
llama_model_loader: - kv  45:                          general.file_type u32              = 12
llama_model_loader: - kv  46:                      quantize.imatrix.file str              = MiniMax-M2.5-GGUF/imatrix_unsloth.gguf
llama_model_loader: - kv  47:                   quantize.imatrix.dataset str              = unsloth_calibration_MiniMax-M2.5.txt
llama_model_loader: - kv  48:             quantize.imatrix.entries_count u32              = 496
llama_model_loader: - kv  49:              quantize.imatrix.chunks_count u32              = 81
llama_model_loader: - kv  50:                                   split.no u16              = 0
llama_model_loader: - kv  51:                        split.tensors.count i32              = 809
llama_model_loader: - kv  52:                                split.count u16              = 4
llama_model_loader: - type  f32:  373 tensors
llama_model_loader: - type q3_K:  173 tensors
llama_model_loader: - type q4_K:  232 tensors
llama_model_loader: - type q5_K:   20 tensors
llama_model_loader: - type q6_K:   11 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q3_K - Medium
print_info: file size   = 94.33 GiB (3.54 BPW)
load: 0 unused tokens
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
load: printing all EOG tokens:
load:   - 200004 ('<fim_pad>')
load:   - 200005 ('<reponame>')
load:   - 200020 ('[e~[')
load: special tokens cache size = 54
load: token to piece cache size = 1.3355 MB
print_info: arch                  = minimax-m2
print_info: vocab_only            = 0
print_info: no_alloc              = 0
print_info: n_ctx_train           = 196608
print_info: n_embd                = 3072
print_info: n_embd_inp            = 3072
print_info: n_layer               = 62
print_info: n_head                = 48
print_info: n_head_kv             = 8
print_info: n_rot                 = 64
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                 = 6
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                  = 1536
print_info: n_expert              = 256
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       = 5000000.0
print_info: freq_scale_train      = 1
print_info: n_ctx_orig_yarn       = 196608
print_info: rope_yarn_log_mul     = 0.0000
print_info: rope_finetuned        = unknown
print_info: model type            = 230B.A10B
print_info: model params          = 228.69 B
print_info: general.name          = Minimax-M2.5
print_info: vocab type            = BPE
print_info: n_vocab               = 200064
print_info: n_merges              = 199744
print_info: BOS token             = 200034 ']~!b['
print_info: EOS token             = 200020 '[e~['
print_info: UNK token             = 200021 ']!d~['
print_info: PAD token             = 200004 '<fim_pad>'
print_info: LF token              = 10 'Ċ'
print_info: FIM PRE token         = 200001 '<fim_prefix>'
print_info: FIM SUF token         = 200003 '<fim_suffix>'
print_info: FIM MID token         = 200002 '<fim_middle>'
print_info: FIM PAD token         = 200004 '<fim_pad>'
print_info: FIM REP token         = 200005 '<reponame>'
print_info: EOG token             = 200004 '<fim_pad>'
print_info: EOG token             = 200005 '<reponame>'
print_info: EOG token             = 200020 '[e~['
print_info: max token length      = 256
load_tensors: loading model tensors, this can take a while... (mmap = false, direct_io = false)
load_tensors: offloading output layer to GPU
load_tensors: offloading 61 repeating layers to GPU
load_tensors: offloaded 63/63 layers to GPU
load_tensors:      Vulkan0 model buffer size = 96266.43 MiB
load_tensors:  Vulkan_Host model buffer size =   329.70 MiB
....................................................................................................
common_init_result: added <fim_pad> logit bias = -inf
common_init_result: added <reponame> logit bias = -inf
common_init_result: added [e~[ logit bias = -inf
llama_context: constructing llama_context
llama_context: n_seq_max     = 1
llama_context: n_ctx         = 87296
llama_context: n_ctx_seq     = 87296
llama_context: n_batch       = 512
llama_context: n_ubatch      = 256
llama_context: causal_attn   = 1
llama_context: flash_attn    = enabled
llama_context: kv_unified    = false
llama_context: freq_base     = 5000000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (87296) < n_ctx_train (196608) -- the full capacity of the model will not be utilized
llama_context: Vulkan_Host  output buffer size =     0.76 MiB
llama_kv_cache:    Vulkan0 KV buffer size = 11231.69 MiB
llama_kv_cache: size = 11231.69 MiB ( 87296 cells,  62 layers,  1/1 seqs), K (q8_0): 5615.84 MiB, V (q8_0): 5615.84 MiB
llama_kv_cache: attn_rot_k = 1, n_embd_head_k_all = 128
llama_kv_cache: attn_rot_v = 1, n_embd_head_k_all = 128
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:    Vulkan0 compute buffer size =   198.38 MiB
sched_reserve: Vulkan_Host compute buffer size =    91.33 MiB
sched_reserve: graph nodes  = 4719
sched_reserve: graph splits = 2
sched_reserve: reserve took 113.49 ms, sched copies = 1
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
srv    load_model: initializing slots, n_slots = 1
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 = 87296
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: --clear-idle requires --kv-unified, disabling
init: chat template, example_format: ']~b]system
You are a helpful assistant[e~[
]~b]user
Hello[e~[
]~b]ai
Hi there[e~[
]~b]user
How are you?[e~[
]~b]ai
<think>
'
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  log_server_r: done request: OPTIONS /v1/chat/completions 100.109.163.105 200
srv  params_from_: Chat format: peg-native
slot get_availabl: id  0 | 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, 87296 tokens, 8589934592 est)
srv  get_availabl: prompt cache update took 0.14 ms
slot launch_slot_: id  0 | 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  0 | task 0 | processing task, is_child = 0
slot update_slots: id  0 | task 0 | new prompt, n_ctx_slot = 87296, n_keep = 0, task.n_tokens = 2924
slot update_slots: id  0 | task 0 | n_tokens = 0, memory_seq_rm [0, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 512, batch.n_tokens = 512, progress = 0.175103
slot update_slots: id  0 | task 0 | n_tokens = 512, memory_seq_rm [512, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 1024, batch.n_tokens = 512, progress = 0.350205
slot update_slots: id  0 | task 0 | n_tokens = 1024, memory_seq_rm [1024, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 1536, batch.n_tokens = 512, progress = 0.525308
slot update_slots: id  0 | task 0 | n_tokens = 1536, memory_seq_rm [1536, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 2048, batch.n_tokens = 512, progress = 0.700410
slot update_slots: id  0 | task 0 | n_tokens = 2048, memory_seq_rm [2048, end)
slot update_slots: id  0 | task 0 | prompt processing progress, n_tokens = 2560, batch.n_tokens = 512, progress = 0.875513
slot update_slots: id  0 | task 0 | n_tokens = 2560, memory_seq_rm [2560, end)
reasoning-budget: activated, budget=2147483647 tokens
slot init_sampler: id  0 | task 0 | init sampler, took 1.05 ms, tokens: text = 2924, total = 2924
slot update_slots: id  0 | task 0 | prompt processing done, n_tokens = 2924, batch.n_tokens = 364
srv  log_server_r: done request: POST /v1/chat/completions 100.109.163.105 200
reasoning-budget: deactivated (natural end)
slot print_timing: id  0 | task 0 |
prompt eval time =   23663.78 ms /  2924 tokens (    8.09 ms per token,   123.56 tokens per second)
       eval time =    4016.48 ms /   109 tokens (   36.85 ms per token,    27.14 tokens per second)
      total time =   27680.26 ms /  3033 tokens
slot      release: id  0 | task 0 | stop processing: n_tokens = 3032, 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