Eval bug: incoherent results with GLM-4-32B-0414-GGUF on CUDA

[LostRuins]

Name and Version

llama-b7172-bin-win-cuda-12.4-x64 public release

Operating systems

Windows

GGML backends

CUDA

Hardware

RTX 4090

Models

https://huggingface.co/unsloth/GLM-4-32B-0414-GGUF/blob/main/GLM-4-32B-0414-Q2_K.gguf

Problem description & steps to reproduce

Finally I am able reproduce this in llama.cpp itself. Can be reproduced on RTX 4090 and 5090 in latest CUDA build

Instructions:

• Download llama-b7172-bin-win-cuda-12.4-x64 from releases
• Download this model: https://huggingface.co/unsloth/GLM-4-32B-0414-GGUF/blob/main/GLM-4-32B-0414-Q2_K.gguf (multiple quants are affected, but it's important that you are able to do a full offload so I picked a small one)
• Run this CLI: llama-server.exe -ngl 99 -m GLM-4-32B-0414-Q2_K.gguf
• Send "tell me a joke" to the AI. Result is garbled nonsense
• Now exit and relaunch, offload 50 layers instead, llama-server.exe -ngl 50 -m GLM-4-32B-0414-Q2_K.gguf
• Send "tell me a joke" to the AI. Result is coherent.
• The sweet spot seems to be (59 layers = coherent, 60+ layers = incoherent)
• Flash attention must be enabled (it is on by default now, so I did not include it).
• This is an old bug that occurred at least one month ago and probably longer, likely after a FA refactor.
• Works without issue if FA is off or using vulkan backend.

cc: @JohannesGaessler

For additional reference (optional, supplementary information distinct from this issue): LostRuins#1837

First Bad Commit

No response

Relevant log output

C:\test\llama-b7172-bin-win-cuda-12.4-x64>llama-server.exe -ngl 60 -m C:\test\GLM-4-32B-0414-Q2_K.gguf
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 4090 Laptop GPU, compute capability 8.9, VMM: yes
load_backend: loaded CUDA backend from C:\test\llama-b7172-bin-win-cuda-12.4-x64\ggml-cuda.dll
load_backend: loaded RPC backend from C:\test\llama-b7172-bin-win-cuda-12.4-x64\ggml-rpc.dll
load_backend: loaded CPU backend from C:\test\llama-b7172-bin-win-cuda-12.4-x64\ggml-cpu-haswell.dll
main: setting n_parallel = 4 and kv_unified = true (add -kvu to disable this)
←[0mbuild: 7172 (b78db3bd5) with clang version 19.1.5 for x86_64-pc-windows-msvc
system info: n_threads = 24, n_threads_batch = 24, total_threads = 32

system_info: n_threads = 24 (n_threads_batch = 24) / 32 | CUDA : ARCHS = 500,610,700,750,800,860,890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |

init: using 31 threads for HTTP server
start: binding port with default address family
main: loading model
srv    load_model: loading model 'C:\test\GLM-4-32B-0414-Q2_K.gguf'
llama_model_load_from_file_impl: using device CUDA0 (NVIDIA GeForce RTX 4090 Laptop GPU) (0000:01:00.0) - 15046 MiB free
llama_model_loader: loaded meta data with 44 key-value pairs and 613 tensors from C:\test\GLM-4-32B-0414-Q2_K.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              = glm4
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Glm-4-32B-0414
llama_model_loader: - kv   3:                            general.version str              = 0414
llama_model_loader: - kv   4:                           general.basename str              = Glm-4-32B-0414
llama_model_loader: - kv   5:                       general.quantized_by str              = Unsloth
llama_model_loader: - kv   6:                         general.size_label str              = 32B
llama_model_loader: - kv   7:                            general.license str              = mit
llama_model_loader: - kv   8:                           general.repo_url str              = https://huggingface.co/unsloth
llama_model_loader: - kv   9:                   general.base_model.count u32              = 1
llama_model_loader: - kv  10:                  general.base_model.0.name str              = GLM 4 32B 0414
llama_model_loader: - kv  11:               general.base_model.0.version str              = 0414
llama_model_loader: - kv  12:          general.base_model.0.organization str              = THUDM
llama_model_loader: - kv  13:              general.base_model.0.repo_url str              = https://huggingface.co/THUDM/GLM-4-32...
llama_model_loader: - kv  14:                               general.tags arr[str,2]       = ["unsloth", "text-generation"]
llama_model_loader: - kv  15:                          general.languages arr[str,2]       = ["zh", "en"]
llama_model_loader: - kv  16:                           glm4.block_count u32              = 61
llama_model_loader: - kv  17:                        glm4.context_length u32              = 32768
llama_model_loader: - kv  18:                      glm4.embedding_length u32              = 6144
llama_model_loader: - kv  19:                   glm4.feed_forward_length u32              = 23040
llama_model_loader: - kv  20:                  glm4.attention.head_count u32              = 48
llama_model_loader: - kv  21:               glm4.attention.head_count_kv u32              = 2
llama_model_loader: - kv  22:                        glm4.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  23:      glm4.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  24:                  glm4.attention.key_length u32              = 128
llama_model_loader: - kv  25:                glm4.attention.value_length u32              = 128
llama_model_loader: - kv  26:                  glm4.rope.dimension_count u32              = 64
llama_model_loader: - kv  27:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  28:                         tokenizer.ggml.pre str              = glm4
llama_model_loader: - kv  29:                      tokenizer.ggml.tokens arr[str,151552]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  30:                  tokenizer.ggml.token_type arr[i32,151552]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  31:                      tokenizer.ggml.merges arr[str,318088]  = ["─á ─á", "─á ─á─á─á", "─á─á ─á─á", "...
llama_model_loader: - kv  32:                tokenizer.ggml.eos_token_id u32              = 151336
llama_model_loader: - kv  33:            tokenizer.ggml.padding_token_id u32              = 151330
llama_model_loader: - kv  34:                tokenizer.ggml.eot_token_id u32              = 151336
llama_model_loader: - kv  35:            tokenizer.ggml.unknown_token_id u32              = 151329
llama_model_loader: - kv  36:                tokenizer.ggml.bos_token_id u32              = 151329
llama_model_loader: - kv  37:                    tokenizer.chat_template str              = [gMASK]<sop>\n{%- if tools -%}\n<|syste...
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              = GLM-4-32B-0414-GGUF/imatrix_unsloth.dat
llama_model_loader: - kv  41:                   quantize.imatrix.dataset str              = unsloth_calibration_GLM-4-32B-0414.txt
llama_model_loader: - kv  42:             quantize.imatrix.entries_count u32              = 366
llama_model_loader: - kv  43:              quantize.imatrix.chunks_count u32              = 724
llama_model_loader: - type  f32:  245 tensors
llama_model_loader: - type q2_K:  184 tensors
llama_model_loader: - type q3_K:  122 tensors
llama_model_loader: - type q4_K:   61 tensors
llama_model_loader: - type q6_K:    1 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q2_K - Medium
print_info: file size   = 11.44 GiB (3.02 BPW)
load: special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect
←[0mload: printing all EOG tokens:
load:   - 151329 ('<|endoftext|>')
load:   - 151336 ('<|user|>')
load: special tokens cache size = 14
load: token to piece cache size = 0.9710 MB
print_info: arch             = glm4
print_info: vocab_only       = 0
print_info: n_ctx_train      = 32768
print_info: n_embd           = 6144
print_info: n_embd_inp       = 6144
print_info: n_layer          = 61
print_info: n_head           = 48
print_info: n_head_kv        = 2
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            = 24
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-05
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             = 23040
print_info: n_expert         = 0
print_info: n_expert_used    = 0
print_info: n_expert_groups  = 0
print_info: n_group_used     = 0
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 0
print_info: rope scaling     = linear
print_info: freq_base_train  = 10000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 32768
print_info: rope_finetuned   = unknown
print_info: model type       = 32B
print_info: model params     = 32.57 B
print_info: general.name     = Glm-4-32B-0414
print_info: vocab type       = BPE
print_info: n_vocab          = 151552
print_info: n_merges         = 318088
print_info: BOS token        = 151329 '<|endoftext|>'
print_info: EOS token        = 151336 '<|user|>'
print_info: EOT token        = 151336 '<|user|>'
print_info: UNK token        = 151329 '<|endoftext|>'
print_info: PAD token        = 151330 '[MASK]'
print_info: LF token         = 198 '─è'
print_info: EOG token        = 151329 '<|endoftext|>'
print_info: EOG token        = 151336 '<|user|>'
print_info: max token length = 1024
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 60 repeating layers to GPU
load_tensors: offloaded 60/62 layers to GPU
load_tensors:   CPU_Mapped model buffer size =  1195.15 MiB
load_tensors:        CUDA0 model buffer size = 10518.75 MiB
..............................................................................................
llama_context: constructing llama_context
llama_context: n_seq_max     = 4
llama_context: n_ctx         = 4096
llama_context: n_ctx_seq     = 4096
llama_context: n_batch       = 2048
llama_context: n_ubatch      = 512
llama_context: causal_attn   = 1
llama_context: flash_attn    = auto
llama_context: kv_unified    = true
llama_context: freq_base     = 10000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_seq (4096) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
←[0mllama_context:        CPU  output buffer size =     2.31 MiB
llama_kv_cache:        CPU KV buffer size =     4.00 MiB
llama_kv_cache:      CUDA0 KV buffer size =   240.00 MiB
llama_kv_cache: size =  244.00 MiB (  4096 cells,  61 layers,  4/1 seqs), K (f16):  122.00 MiB, V (f16):  122.00 MiB
llama_context: Flash Attention was auto, set to enabled
llama_context:      CUDA0 compute buffer size =  1036.44 MiB
llama_context:  CUDA_Host compute buffer size =    20.01 MiB
llama_context: graph nodes  = 2081
llama_context: graph splits = 16 (with bs=512), 3 (with bs=1)
common_init_from_params: added <|endoftext|> logit bias = -inf
common_init_from_params: added <|user|> 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)
←[0msrv          init: initializing slots, n_slots = 4
slot         init: id  0 | task -1 | new slot, n_ctx = 4096
slot         init: id  1 | task -1 | new slot, n_ctx = 4096
slot         init: id  2 | task -1 | new slot, n_ctx = 4096
slot         init: id  3 | task -1 | new slot, n_ctx = 4096
srv          init: prompt cache is enabled, size limit: 8192 MiB
←[0msrv          init: use `--cache-ram 0` to disable the prompt cache
←[0msrv          init: for more info see https://github.com/ggml-org/llama.cpp/pull/16391
←[0msrv          init: thinking = 0
init: chat template, chat_template: [gMASK]<sop>
{%- if tools -%}
<|system|>
# 可用工具
{% for tool in tools %}
    {%- set function = tool.function if tool.get("function") else tool %}

## {{ function.name }}

{{ function | tojson(indent=4)|string }}
在调用上述函数时，请使用 Json 格式表示调用的参数。
{%- endfor %}
{%- endif -%}

{%- for msg in messages %}
    {%- if msg.role == 'system' %}
<|system|>
{{ msg.content }}
    {%- endif %}
{%- endfor %}

{%- for message in messages if message.role != 'system' %}
    {%- set role = message['role'] %}
    {%- set content = message['content'] %}
    {%- set meta = message.get("metadata", "") %}

    {%- if role == 'user' %}
<|user|>
{{ content }}
    {%- elif role == 'assistant' and not meta %}
<|assistant|>
{{ content }}
    {%- elif role == 'assistant' and meta %}
<|assistant|>{{ meta }}
{{ content }}
    {%- elif role == 'observation' %}
<|observation|>
{{ content }}
    {%- endif %}
{%- endfor %}
{% if add_generation_prompt %}<|assistant|>{% endif %}, example_format: '[gMASK]<sop><|system|>
You are a helpful assistant<|user|>
Hello<|assistant|>
Hi there<|user|>
How are you?<|assistant|>'
main: model loaded
main: server is listening on http://127.0.0.1:8080
main: starting the main loop...
srv  update_slots: all slots are idle
srv  update_slots: all slots are idle
srv  log_server_r: request: GET /slots 127.0.0.1 200
srv  params_from_: Chat format: Content-only
slot get_availabl: id  3 | task -1 | selected slot by LRU, t_last = -1
slot launch_slot_: id  3 | task -1 | sampler chain: logits -> logit-bias -> penalties -> dry -> top-n-sigma -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
slot launch_slot_: id  3 | task 1 | processing task
slot update_slots: id  3 | task 1 | new prompt, n_ctx_slot = 4096, n_keep = 0, task.n_tokens = 9
slot update_slots: id  3 | task 1 | n_tokens = 0, memory_seq_rm [0, end)
slot update_slots: id  3 | task 1 | prompt processing progress, n_tokens = 9, batch.n_tokens = 9, progress = 1.000000
slot update_slots: id  3 | task 1 | prompt done, n_tokens = 9, batch.n_tokens = 9
srv  log_server_r: request: GET /props 127.0.0.1 200
srv  log_server_r: request: GET /slots 127.0.0.1 200
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 200
srv          stop: cancel task, id_task = 1
←[0mslot      release: id  3 | task 1 | stop processing: n_tokens = 52, truncated = 0
srv  update_slots: all slots are idle

[CISC]

Likely related to #16945 however there it seemed to be reverse (fully offloaded was ok).

[LostRuins]

GLM4 is not a MoE though, so may not be the same thing. I'm just glad i can finally repro it here.

[CISC]

GLM4 is not a MoE though, so may not be the same thing. I'm just glad i can finally repro it here.

Yeah, though I don't think it matters, it was just easier to trigger with MoE.

[pwilkin]

Tactical lurker comment.

[pwilkin]

So I managed to sort-of-reproduce this, I had to explicitly set offload to one card:

llama-server: /devel/tools/llama.cpp/src/llama-sampling.cpp:660: void llama_sampler_dist_apply(llama_sampler*, llama_token_data_array*): Assertion `found' failed.

[pwilkin]

So first of all, there are tiny MATMUL differences between kernels:

ggml_debug:                   node_3 = (f32)    MUL_MAT(blk.0.attn_q.weight{6144, 6144, 1, 1}, attn_norm-0{6144, 10, 1, 1}}) = {6144, 10, 1, 1}
                                     [
                                      [
                                       [      4.1491,       2.7689,      -2.9154, ...,      -3.1288,       0.2312,      -0.6199],
                                       [     -3.8194,      -3.2223,       4.2295, ...,       5.8112,       2.0654,       0.2569],
                                       [     -2.0825,       2.1551,      -2.1050, ...,       2.3807,       2.6710,       1.3906],
                                       ..., 
                                       [     -0.4999,       1.7304,      -3.9699, ...,       0.5508,       2.7632,       0.5763],
                                       [     -3.8194,      -3.2223,       4.2295, ...,       5.8112,       2.0654,       0.2569],
                                       [     -2.8283,      -1.6989,       1.6393, ...,      -2.4234,      -2.5988,       2.1787],
                                      ],
                                     ]
                                     sum = -3126.369629

(this is the 5060)

ggml_debug:                   node_3 = (f32)    MUL_MAT(blk.0.attn_q.weight{6144, 6144, 1, 1}, attn_norm-0{6144, 10, 1, 1}}) = {6144, 10, 1, 1}
                                     [
                                      [
                                       [      4.1491,       2.7689,      -2.9154, ...,      -3.1287,       0.2312,      -0.6199],
                                       [     -3.8194,      -3.2223,       4.2295, ...,       5.8112,       2.0654,       0.2569],
                                       [     -2.0825,       2.1551,      -2.1050, ...,       2.3807,       2.6710,       1.3906],
                                       ..., 
                                       [     -0.4999,       1.7304,      -3.9699, ...,       0.5508,       2.7632,       0.5763],
                                       [     -3.8194,      -3.2223,       4.2295, ...,       5.8112,       2.0654,       0.2569],
                                       [     -2.8283,      -1.6989,       1.6393, ...,      -2.4234,      -2.5988,       2.1787],
                                      ],
                                     ]
                                     sum = -3126.367676

(this is the 3080)

Second of all, SWIGLU looking mighty suspicious right now:

ggml_debug:             ffn_swiglu-2 = (f32)     SWIGLU(ffn_up-2{46080, 10, 1, 1}, }) = {23040, 10, 1, 1}
                                     [
                                      [
                                       [     -0.1177,      -0.0003,       0.0043, ...,      -0.1912,      -0.9875,      -0.0369],
                                       [      0.1196,      -0.0309,       0.4960, ...,       0.1146,      -0.1472,      -0.0171],
                                       [      0.2936,       0.0051,       0.2135, ...,       0.3188,      -0.6223,      -0.5825],
                                       ..., 
                                       [      0.3347,      -0.0001,      -0.1918, ...,       0.1387,      -0.7570,      -0.0041],
                                       [      0.1331,      -0.0280,       0.2235, ...,       0.0919,       0.1071,      -0.0272],
                                       [      0.0482,       0.1285,       0.3361, ...,       0.0560,       0.0816,      -0.0628],
                                      ],
                                     ]
                                     sum = -104.895187

vs

ggml_debug:             ffn_swiglu-2 = (f32)     SWIGLU(ffn_up-2{46080, 10, 1, 1}, }) = {23040, 10, 1, 1}
                                     [
                                      [
                                       [     -0.1177,      -0.0003,       0.0044, ...,      -0.1909,      -0.9873,      -0.0369],
                                       [      0.1223,      -0.0311,       0.4957, ...,       0.1141,      -0.1470,      -0.0170],
                                       [      0.2938,       0.0052,       0.2184, ...,       0.3225,      -0.6250,      -0.5801],
                                       ..., 
                                       [      0.3328,      -0.0001,      -0.1899, ...,       0.1365,      -0.7499,      -0.0047],
                                       [      0.1366,      -0.0261,       0.2275, ...,       0.0928,       0.1125,      -0.0263],
                                       [      0.0488,       0.1275,       0.3352, ...,       0.0518,       0.0781,      -0.0649],
                                      ],
                                     ]
                                     sum = -88.741501

[JohannesGaessler]

I already know the exact issue and how to fix it.