musa: enable MMA

[yeahdongcn]

Make sure to read the contributing guidelines before submitting a PR

This PR enables muBLAS and MMA support on MUSA (QY2).
Edit:

I will rebase once #13144 is merged into master.

Done

Testing Done

root@2f2c227f02d8:/ws# ./build/bin/llama-cli -m /models/models/Ling-lite/Ling-lite.Q4_K_M.gguf -ngl 999 -fa
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 MUSA devices:
  Device 0: MTT S4000, compute capability 2.2, VMM: yes
build: 5202 (fd00e2d2) with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 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 MUSA0 (MTT S4000) - 43580 MiB free
llama_model_loader: loaded meta data with 44 key-value pairs and 367 tensors from /models/models/Ling-lite/Ling-lite.Q4_K_M.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              = bailingmoe
llama_model_loader: - kv   1:                               general.type str              = model
llama_model_loader: - kv   2:                               general.name str              = Ling Lite
llama_model_loader: - kv   3:                         general.size_label str              = 64x1.5B
llama_model_loader: - kv   4:                            general.license str              = mit
llama_model_loader: - kv   5:                               general.tags arr[str,1]       = ["text-generation"]
llama_model_loader: - kv   6:                     bailingmoe.block_count u32              = 28
llama_model_loader: - kv   7:                  bailingmoe.context_length u32              = 16384
llama_model_loader: - kv   8:                bailingmoe.embedding_length u32              = 2048
llama_model_loader: - kv   9:             bailingmoe.feed_forward_length u32              = 5632
llama_model_loader: - kv  10:            bailingmoe.attention.head_count u32              = 16
llama_model_loader: - kv  11:         bailingmoe.attention.head_count_kv u32              = 4
llama_model_loader: - kv  12:                  bailingmoe.rope.freq_base f32              = 600000.000000
llama_model_loader: - kv  13: bailingmoe.attention.layer_norm_rms_epsilon f32              = 0.000001
llama_model_loader: - kv  14:               bailingmoe.expert_used_count u32              = 6
llama_model_loader: - kv  15:            bailingmoe.rope.dimension_count u32              = 128
llama_model_loader: - kv  16:               bailingmoe.rope.scaling.type str              = none
llama_model_loader: - kv  17:       bailingmoe.leading_dense_block_count u32              = 0
llama_model_loader: - kv  18:                      bailingmoe.vocab_size u32              = 126464
llama_model_loader: - kv  19:      bailingmoe.expert_feed_forward_length u32              = 1408
llama_model_loader: - kv  20:            bailingmoe.expert_weights_scale f32              = 1.000000
llama_model_loader: - kv  21:                    bailingmoe.expert_count u32              = 64
llama_model_loader: - kv  22:             bailingmoe.expert_shared_count u32              = 2
llama_model_loader: - kv  23:             bailingmoe.expert_weights_norm bool             = true
llama_model_loader: - kv  24:                       tokenizer.ggml.model str              = gpt2
llama_model_loader: - kv  25:                         tokenizer.ggml.pre str              = bailingmoe
llama_model_loader: - kv  26:                      tokenizer.ggml.tokens arr[str,126464]  = ["!", "\"", "#", "$", "%", "&", "'", ...
llama_model_loader: - kv  27:                  tokenizer.ggml.token_type arr[i32,126464]  = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  28:                      tokenizer.ggml.merges arr[str,125824]  = ["Ġ Ġ", "Ġ t", "i n", "Ġ a", "h e...
llama_model_loader: - kv  29:                tokenizer.ggml.bos_token_id u32              = 126080
llama_model_loader: - kv  30:                tokenizer.ggml.eos_token_id u32              = 126081
llama_model_loader: - kv  31:            tokenizer.ggml.padding_token_id u32              = 126081
llama_model_loader: - kv  32:               tokenizer.ggml.add_bos_token bool             = false
llama_model_loader: - kv  33:               tokenizer.ggml.add_eos_token bool             = false
llama_model_loader: - kv  34:                    tokenizer.chat_template str              = {% for message in messages %}{% set r...
llama_model_loader: - kv  35:               general.quantization_version u32              = 2
llama_model_loader: - kv  36:                          general.file_type u32              = 15
llama_model_loader: - kv  37:                                general.url str              = https://huggingface.co/mradermacher/L...
llama_model_loader: - kv  38:              mradermacher.quantize_version str              = 2
llama_model_loader: - kv  39:                  mradermacher.quantized_by str              = mradermacher
llama_model_loader: - kv  40:                  mradermacher.quantized_at str              = 2025-03-31T05:37:59+02:00
llama_model_loader: - kv  41:                  mradermacher.quantized_on str              = kaos
llama_model_loader: - kv  42:                         general.source.url str              = https://huggingface.co/inclusionAI/Li...
llama_model_loader: - kv  43:                  mradermacher.convert_type str              = hf
llama_model_loader: - type  f32:   85 tensors
llama_model_loader: - type q5_0:   14 tensors
llama_model_loader: - type q8_0:   14 tensors
llama_model_loader: - type q4_K:  225 tensors
llama_model_loader: - type q6_K:   29 tensors
print_info: file format = GGUF V3 (latest)
print_info: file type   = Q4_K - Medium
print_info: file size   = 10.40 GiB (5.32 BPW) 
load: special tokens cache size = 266
load: token to piece cache size = 0.8056 MB
print_info: arch             = bailingmoe
print_info: vocab_only       = 0
print_info: n_ctx_train      = 16384
print_info: n_embd           = 2048
print_info: n_layer          = 28
print_info: n_head           = 16
print_info: n_head_kv        = 4
print_info: n_rot            = 128
print_info: n_swa            = 0
print_info: n_swa_pattern    = 1
print_info: n_embd_head_k    = 128
print_info: n_embd_head_v    = 128
print_info: n_gqa            = 4
print_info: n_embd_k_gqa     = 512
print_info: n_embd_v_gqa     = 512
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             = 5632
print_info: n_expert         = 64
print_info: n_expert_used    = 6
print_info: causal attn      = 1
print_info: pooling type     = 0
print_info: rope type        = 0
print_info: rope scaling     = none
print_info: freq_base_train  = 600000.0
print_info: freq_scale_train = 1
print_info: n_ctx_orig_yarn  = 16384
print_info: rope_finetuned   = unknown
print_info: ssm_d_conv       = 0
print_info: ssm_d_inner      = 0
print_info: ssm_d_state      = 0
print_info: ssm_dt_rank      = 0
print_info: ssm_dt_b_c_rms   = 0
print_info: model type       = 16B
print_info: model params     = 16.80 B
print_info: general.name     = Ling Lite
print_info: n_layer_dense_lead   = 0
print_info: n_ff_exp             = 1408
print_info: n_expert_shared      = 2
print_info: expert_weights_scale = 1.0
print_info: expert_weights_norm  = 1
print_info: vocab type       = BPE
print_info: n_vocab          = 126464
print_info: n_merges         = 125824
print_info: BOS token        = 126080 '<|startoftext|>'
print_info: EOS token        = 126081 '<|endoftext|>'
print_info: EOT token        = 126081 '<|endoftext|>'
print_info: PAD token        = 126081 '<|endoftext|>'
print_info: LF token         = 198 'Ċ'
print_info: EOG token        = 126081 '<|endoftext|>'
print_info: max token length = 154
load_tensors: loading model tensors, this can take a while... (mmap = true)
load_tensors: offloading 28 repeating layers to GPU
load_tensors: offloading output layer to GPU
load_tensors: offloaded 29/29 layers to GPU
load_tensors:        MUSA0 model buffer size = 10513.91 MiB
load_tensors:   CPU_Mapped model buffer size =   138.94 MiB
......................................................................................
llama_context: constructing llama_context
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    = 1
llama_context: freq_base     = 600000.0
llama_context: freq_scale    = 1
llama_context: n_ctx_per_seq (4096) < n_ctx_train (16384) -- the full capacity of the model will not be utilized
llama_context:  MUSA_Host  output buffer size =     0.48 MiB
init: kv_size = 4096, offload = 1, type_k = 'f16', type_v = 'f16', n_layer = 28, can_shift = 1
init:      MUSA0 KV buffer size =   224.00 MiB
llama_context: KV self size  =  224.00 MiB, K (f16):  112.00 MiB, V (f16):  112.00 MiB
llama_context:      MUSA0 compute buffer size =   259.00 MiB
llama_context:  MUSA_Host compute buffer size =    20.01 MiB
llama_context: graph nodes  = 1659
llama_context: graph splits = 58
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 = 6
main: chat template is available, enabling conversation mode (disable it with -no-cnv)
main: chat template example:
<role>SYSTEM</role>You are a helpful assistant<role>HUMAN</role>Hello<role>ASSISTANT</role>Hi there<role>HUMAN</role>How are you?<role>ASSISTANT</role>

system_info: n_threads = 6 (n_threads_batch = 6) / 12 | MUSA : PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | AARCH64_REPACK = 1 | 

main: interactive mode on.
sampler seed: 1499548142
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-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
Hello! How can I assist you today?

> Do you know math?
Yes, I have knowledge of various mathematical concepts and can help with solving problems, explaining formulas, and providing information on different mathematical topics. If you have a specific math question or need help with something, feel free to ask!

> 
llama_perf_sampler_print:    sampling time =       3.30 ms /    60 runs   (    0.05 ms per token, 18192.84 tokens per second)
llama_perf_context_print:        load time =    2193.15 ms
llama_perf_context_print: prompt eval time =    1496.69 ms /    25 tokens (   59.87 ms per token,    16.70 tokens per second)
llama_perf_context_print:        eval time =    2534.62 ms /    54 runs   (   46.94 ms per token,    21.30 tokens per second)
llama_perf_context_print:       total time =   13538.23 ms /    79 tokens
Interrupted by user
root@2f2c227f02d8:/ws# 

[yeahdongcn]

So far, I can only get it working with -fa. Without -fa, I encounter either garbled characters in the LLM replies or repeated GGGGGGG....

@JohannesGaessler @slaren Could you please share some tips on how to debug this kind of issue? I'd appreciate it!

[JohannesGaessler]

Why are you manually allocating and deallocating memory instead of using ggml_cuda_pool_alloc? Batched FP16 GEMM is used for attention without FlashAttention so most likely this is where the bug is. I don't remember what the synchronization behavior of cudaFree is but if it's done asynchronously from the kernel executions that would explain why you get incorrect results.

[JohannesGaessler]

Could you please share some tips on how to debug this kind of issue?

Run test-backend-ops -o MUL_MAT.