Bug: Garbage outputs on Vulkan backend since #10301 (possible NaN issue?)

[stduhpf]

What happened?

Since commit b3e5859 (PR #10301), The vulkan backend tends to glitch out after a few tokens, this happened to me a few times with llama-3 8B, and happens consistently with qwen2.5-Coder 0.5B. My guess is that there must be invalid values like NaN or Inf appearing somewhere during computation of softmax.

How to reproduce.

Command:

.\build\bin\Release\llama-cli.exe -m E:\Downloads\Qwen2.5-Coder-0.5B.f16.gguf -ngl 99 -t 6 -tb 12 -p "Hello, I'm a" --seed 0 -n 512

Output:

Hello, I'm a little bit of a beginner with C++ and I have a question regarding the std::vector. I am trying to add an element to the end of the vector, but I am getting a segmentation fault error. Could you help me with this? Sure, I can help with that. To add an element to the end of a vector in C++,GGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGG

Expected output (commit 557924f)

Hello, I'm a little bit of a beginner with C++ and I have a question regarding the std::vector. I am trying to add an element to the end of the vector, but I am getting a segmentation fault error. Could you help me with this? Sure, I can help with that. To add an element to the end of a vector in C++, you can use the push_back() function. The push_back() function takes a single argument which is the element you want to add to the vector. Here is an example:

vector<int> myVector;
myVector.push_back(5);

In the above code, myVector is a reference to a vector of integers. The push_back() function is used to add an element to the end of the vector. The element is specified after the push_back() function call. If you want to add multiple elements at once, you can separate them with commas.

Please let me know if you have any other questions. [end of text]

Git bisect results

b3e585988fc65d3a8083c6d94dfc0629f9ce226d is the first bad commit
commit b3e585988fc65d3a8083c6d94dfc0629f9ce226d (tag: b4128)
Author: Jeff Bolz <jbolz@nvidia.com>
Date:   Tue Nov 19 01:25:17 2024 -0600

    vulkan: Optimize soft_max (#10301)

    * vulkan: Optimize soft_max

    Large soft_max could already saturate memory, but small/medium sizes were
    pretty slow. The bulk of the gains for them comes from using a smaller
    workgroup size, and making the workgroup size match the subgroup size also
    makes the barriers much cheaper.

    Cache some values in locals to avoid refetching/recomputing. And stamp
    out a few "template instantiations" so smaller cases will fully unroll.

    Add a missing early return for OOB rows. This happens when there are more
    than 512 rows and the dispatch is 512 x H.

    * vulkan: Further soft_max optimizations

    Restore the workgroup size of 512 case, use it for >1024.

    Use unrollable loops for more iteration counts.

 ggml/src/ggml-vulkan/ggml-vulkan.cpp              |  13 ++-
 ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp | 112 +++++++++++++++++-----
 tests/test-backend-ops.cpp                        |   8 ++
 3 files changed, 106 insertions(+), 27 deletions(-)

Full logs

ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon RX 5700 XT (AMD proprietary driver) | uma: 0 | fp16: 1 | warp size: 64
build: 4128 (b3e585988) with MSVC 19.41.34120.0 for x64
main: llama backend init
main: load the model and apply lora adapter, if any
llama_load_model_from_file: using device Vulkan0 (AMD Radeon RX 5700 XT) - 8176 MiB free
llama_model_loader: loaded meta data with 40 key-value pairs and 290 tensors from E:\Downloads\Qwen2.5-Coder-0.5B.f16.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              = qwen2
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Qwen2.5 Coder 0.5B
llama_model_loader: - kv   3:                           general.basename str              = Qwen2.5-Coder
llama_model_loader: - kv   4:                         general.size_label str              = 0.5B
llama_model_loader: - kv   5:                            general.license str              = apache-2.0
llama_model_loader: - kv   6:                       general.license.link str              = https://huggingface.co/Qwen/Qwen2.5-C...
llama_model_loader: - kv   7:                   general.base_model.count u32              = 1
llama_model_loader: - kv   8:                  general.base_model.0.name str              = Qwen2.5 0.5B
llama_model_loader: - kv   9:          general.base_model.0.organization str              = Qwen
llama_model_loader: - kv  10:              general.base_model.0.repo_url str              = https://huggingface.co/Qwen/Qwen2.5-0.5B
llama_model_loader: - kv  11:                               general.tags arr[str,5]       = ["code", "qwen", "qwen-coder", "codeq...
llama_model_loader: - kv  12:                          general.languages arr[str,1]       = ["en"]
llama_model_loader: - kv  13:                          qwen2.block_count u32              = 24
llama_model_loader: - kv  14:                       qwen2.context_length u32              = 32768
llama_model_loader: - kv  15:                     qwen2.embedding_length u32              = 896
llama_model_loader: - kv  16:                  qwen2.feed_forward_length u32              = 4864
llama_model_loader: - kv  17:                 qwen2.attention.head_count u32              = 14
llama_model_loader: - kv  18:              qwen2.attention.head_count_kv u32              = 2
llama_model_loader: - kv  19:                       qwen2.rope.freq_base f32              = 1000000.000000
llama_model_loader: - kv  20:     qwen2.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  21:                          general.file_type u32              = 1
llama_model_loader: - kv  22:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  23:                         tokenizer.ggml.pre str              = qwen2
llama_model_loader: - kv  24:                      tokenizer.ggml.tokens arr[str,151936]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  25:                  tokenizer.ggml.token_type arr[i32,151936]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  26:                      tokenizer.ggml.merges arr[str,151387]  = ["Ġ Ġ", "ĠĠ ĠĠ", "i n", "Ġ t",...
llama_model_loader: - kv  27:                tokenizer.ggml.eos_token_id u32              = 151645
llama_model_loader: - kv  28:            tokenizer.ggml.padding_token_id u32              = 151643
llama_model_loader: - kv  29:                tokenizer.ggml.bos_token_id u32              = 151643
llama_model_loader: - kv  30:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  31:                    tokenizer.chat_template str              = {%- if tools %}\n    {{- '<|im_start|>...
llama_model_loader: - kv  32:               general.quantization_version u32              = 2
llama_model_loader: - kv  33:                                general.url str              = https://huggingface.co/mradermacher/Q...
llama_model_loader: - kv  34:              mradermacher.quantize_version str              = 2
llama_model_loader: - kv  35:                  mradermacher.quantized_by str              = mradermacher
llama_model_loader: - kv  36:                  mradermacher.quantized_at str              = 2024-11-12T06:29:06+01:00
llama_model_loader: - kv  37:                  mradermacher.quantized_on str              = db2
llama_model_loader: - kv  38:                         general.source.url str              = https://huggingface.co/Qwen/Qwen2.5-C...
llama_model_loader: - kv  39:                  mradermacher.convert_type str              = hf
llama_model_loader: - type  f32:  121 tensors
llama_model_loader: - type  f16:  169 tensors
llm_load_vocab: special tokens cache size = 22
llm_load_vocab: token to piece cache size = 0.9310 MB
llm_load_print_meta: format           = GGUF V3 (latest)
llm_load_print_meta: arch             = qwen2
llm_load_print_meta: vocab type       = BPE
llm_load_print_meta: n_vocab          = 151936
llm_load_print_meta: n_merges         = 151387
llm_load_print_meta: vocab_only       = 0
llm_load_print_meta: n_ctx_train      = 32768
llm_load_print_meta: n_embd           = 896
llm_load_print_meta: n_layer          = 24
llm_load_print_meta: n_head           = 14
llm_load_print_meta: n_head_kv        = 2
llm_load_print_meta: n_rot            = 64
llm_load_print_meta: n_swa            = 0
llm_load_print_meta: n_embd_head_k    = 64
llm_load_print_meta: n_embd_head_v    = 64
llm_load_print_meta: n_gqa            = 7
llm_load_print_meta: n_embd_k_gqa     = 128
llm_load_print_meta: n_embd_v_gqa     = 128
llm_load_print_meta: f_norm_eps       = 0.0e+00
llm_load_print_meta: f_norm_rms_eps   = 1.0e-06
llm_load_print_meta: f_clamp_kqv      = 0.0e+00
llm_load_print_meta: f_max_alibi_bias = 0.0e+00
llm_load_print_meta: f_logit_scale    = 0.0e+00
llm_load_print_meta: n_ff             = 4864
llm_load_print_meta: n_expert         = 0
llm_load_print_meta: n_expert_used    = 0
llm_load_print_meta: causal attn      = 1
llm_load_print_meta: pooling type     = 0
llm_load_print_meta: rope type        = 2
llm_load_print_meta: rope scaling     = linear
llm_load_print_meta: freq_base_train  = 1000000.0
llm_load_print_meta: freq_scale_train = 1
llm_load_print_meta: n_ctx_orig_yarn  = 32768
llm_load_print_meta: rope_finetuned   = unknown
llm_load_print_meta: ssm_d_conv       = 0
llm_load_print_meta: ssm_d_inner      = 0
llm_load_print_meta: ssm_d_state      = 0
llm_load_print_meta: ssm_dt_rank      = 0
llm_load_print_meta: ssm_dt_b_c_rms   = 0
llm_load_print_meta: model type       = 1B
llm_load_print_meta: model ftype      = F16
llm_load_print_meta: model params     = 494.03 M
llm_load_print_meta: model size       = 942.43 MiB (16.00 BPW)
llm_load_print_meta: general.name     = Qwen2.5 Coder 0.5B
llm_load_print_meta: BOS token        = 151643 '<|endoftext|>'
llm_load_print_meta: EOS token        = 151645 '<|im_end|>'
llm_load_print_meta: EOT token        = 151645 '<|im_end|>'
llm_load_print_meta: PAD token        = 151643 '<|endoftext|>'
llm_load_print_meta: LF token         = 148848 'ÄĬ'
llm_load_print_meta: FIM PRE token    = 151659 '<|fim_prefix|>'
llm_load_print_meta: FIM SUF token    = 151661 '<|fim_suffix|>'
llm_load_print_meta: FIM MID token    = 151660 '<|fim_middle|>'
llm_load_print_meta: FIM PAD token    = 151662 '<|fim_pad|>'
llm_load_print_meta: FIM REP token    = 151663 '<|repo_name|>'
llm_load_print_meta: FIM SEP token    = 151664 '<|file_sep|>'
llm_load_print_meta: EOG token        = 151643 '<|endoftext|>'
llm_load_print_meta: EOG token        = 151645 '<|im_end|>'
llm_load_print_meta: EOG token        = 151662 '<|fim_pad|>'
llm_load_print_meta: EOG token        = 151663 '<|repo_name|>'
llm_load_print_meta: EOG token        = 151664 '<|file_sep|>'
llm_load_print_meta: max token length = 256
ggml_vulkan: Compiling shaders..............................Done!
llm_load_tensors: offloading 24 repeating layers to GPU
llm_load_tensors: offloading output layer to GPU
llm_load_tensors: offloaded 25/25 layers to GPU
llm_load_tensors:      Vulkan0 model buffer size =   942.43 MiB
llm_load_tensors:   CPU_Mapped model buffer size =   259.66 MiB
...........................................................
llama_new_context_with_model: n_seq_max     = 1
llama_new_context_with_model: n_ctx         = 4096
llama_new_context_with_model: n_ctx_per_seq = 4096
llama_new_context_with_model: n_batch       = 2048
llama_new_context_with_model: n_ubatch      = 512
llama_new_context_with_model: flash_attn    = 0
llama_new_context_with_model: freq_base     = 1000000.0
llama_new_context_with_model: freq_scale    = 1
llama_new_context_with_model: n_ctx_per_seq (4096) < n_ctx_train (32768) -- the full capacity of the model will not be utilized
llama_kv_cache_init:    Vulkan0 KV buffer size =    48.00 MiB
llama_new_context_with_model: KV self size  =   48.00 MiB, K (f16):   24.00 MiB, V (f16):   24.00 MiB
llama_new_context_with_model: Vulkan_Host  output buffer size =     0.58 MiB
llama_new_context_with_model:    Vulkan0 compute buffer size =   298.50 MiB
llama_new_context_with_model: Vulkan_Host compute buffer size =     9.76 MiB
llama_new_context_with_model: graph nodes  = 846
llama_new_context_with_model: graph splits = 2
common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
main: llama threadpool init, n_threads = 6

system_info: n_threads = 6 (n_threads_batch = 12) / 24 | AVX = 1 | AVX_VNNI = 0 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | AMX_INT8 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | RISCV_VECT = 0 | WASM_SIMD = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 |

sampler seed: 0
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 = -1
        top_k = 40, top_p = 0.950, min_p = 0.050, xtc_probability = 0.000, xtc_threshold = 0.100, typical_p = 1.000, temp = 0.800
        mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampler chain: logits -> logit-bias -> penalties -> dry -> top-k -> typical -> top-p -> min-p -> xtc -> temp-ext -> dist
generate: n_ctx = 4096, n_batch = 2048, n_predict = 512, n_keep = 0

Hello, I'm a little bit of a beginner with C++ and I have a question regarding the `std::vector`. I am trying to add an element to the end of the vector, but I am getting a segmentation fault error. Could you help me with this? Sure, I can help with that. To add an element to the end of a vector in C++,GGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGGG

llama_perf_sampler_print:    sampling time =      46.22 ms /   517 runs   (    0.09 ms per token, 11186.60 tokens per second)
llama_perf_context_print:        load time =    1177.49 ms
llama_perf_context_print: prompt eval time =      50.35 ms /     5 tokens (   10.07 ms per token,    99.30 tokens per second)
llama_perf_context_print:        eval time =    2678.84 ms /   511 runs   (    5.24 ms per token,   190.75 tokens per second)
llama_perf_context_print:       total time =    2830.83 ms /   516 tokens

Name and Version

.\build\bin\Release\llama-cli.exe --version
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon RX 5700 XT (AMD proprietary driver) | uma: 0 | fp16: 1 | warp size: 64
version: 4128 (b3e5859)
built with MSVC 19.41.34120.0 for x64

What operating system are you seeing the problem on?

Windows

[jeffbolznv]

I think this is caused by a missing break in the first loop. Let me do some quick perf tests and I'll make a PR.