musa: enable building fat binaries, enable unified memory, and disable Flash Attention on QY1 (MTT S80)

[yeahdongcn]

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

This PR enables building fat binaries for both QY1 (mp_21, MTT S80) and QY2 (mp_22, MTT S4000), and enables unified memory (#8035). However, due to a known issue when compiling Flash Attention on QY1, it has been explicitly disabled.

Testing done

• make GGML_MUSA=1 -> passed

• ran tinyllama model on MTT S80 and MTT S4000 (w/wo GGML_CUDA_ENABLE_UNIFIED_MEMORY=1) -> passed

  root@e9c2e4c9121b:/ws# ./llama-cli -ngl 999
  build: 3783 (df79623d) with clang version 14.0.0 ([email protected]:sw/mtcc.git a0d2558b470b4b21e99af11c036fab6801e57075) for x86_64-unknown-linux-gnu
  main: llama backend init
  main: load the model and apply lora adapter, if any
  llama_model_loader: loaded meta data with 23 key-value pairs and 201 tensors from models/7B/ggml-model-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              = llama
  llama_model_loader: - kv   1:                               general.name str              = TinyLlama
  llama_model_loader: - kv   2:                       llama.context_length u32              = 2048
  llama_model_loader: - kv   3:                     llama.embedding_length u32              = 2048
  llama_model_loader: - kv   4:                          llama.block_count u32              = 22
  llama_model_loader: - kv   5:                  llama.feed_forward_length u32              = 5632
  llama_model_loader: - kv   6:                 llama.rope.dimension_count u32              = 64
  llama_model_loader: - kv   7:                 llama.attention.head_count u32              = 32
  llama_model_loader: - kv   8:              llama.attention.head_count_kv u32              = 4
  llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32              = 0.000010
  llama_model_loader: - kv  10:                       llama.rope.freq_base f32              = 10000.000000
  llama_model_loader: - kv  11:                          general.file_type u32              = 2
  llama_model_loader: - kv  12:                       tokenizer.ggml.model str              = llama
  llama_model_loader: - kv  13:                      tokenizer.ggml.tokens arr[str,32000]   = ["<unk>", "<s>", "</s>", "<0x00>", "<...
  llama_model_loader: - kv  14:                      tokenizer.ggml.scores arr[f32,32000]   = [0.000000, 0.000000, 0.000000, 0.0000...
  llama_model_loader: - kv  15:                  tokenizer.ggml.token_type arr[i32,32000]   = [2, 3, 3, 6, 6, 6, 6, 6, 6, 6, 6, 6, ...
  llama_model_loader: - kv  16:                      tokenizer.ggml.merges arr[str,61249]   = ["▁ t", "e r", "i n", "▁ a", "e n...
  llama_model_loader: - kv  17:                tokenizer.ggml.bos_token_id u32              = 1
  llama_model_loader: - kv  18:                tokenizer.ggml.eos_token_id u32              = 2
  llama_model_loader: - kv  19:            tokenizer.ggml.unknown_token_id u32              = 0
  llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 2
  llama_model_loader: - kv  21:                    tokenizer.chat_template str              = {% for message in messages %}\n{% if m...
  llama_model_loader: - kv  22:               general.quantization_version u32              = 2
  llama_model_loader: - type  f32:   45 tensors
  llama_model_loader: - type q4_0:  155 tensors
  llama_model_loader: - type q6_K:    1 tensors
  llm_load_vocab: special tokens cache size = 3
  llm_load_vocab: token to piece cache size = 0.1684 MB
  llm_load_print_meta: format           = GGUF V3 (latest)
  llm_load_print_meta: arch             = llama
  llm_load_print_meta: vocab type       = SPM
  llm_load_print_meta: n_vocab          = 32000
  llm_load_print_meta: n_merges         = 0
  llm_load_print_meta: vocab_only       = 0
  llm_load_print_meta: n_ctx_train      = 2048
  llm_load_print_meta: n_embd           = 2048
  llm_load_print_meta: n_layer          = 22
  llm_load_print_meta: n_head           = 32
  llm_load_print_meta: n_head_kv        = 4
  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            = 8
  llm_load_print_meta: n_embd_k_gqa     = 256
  llm_load_print_meta: n_embd_v_gqa     = 256
  llm_load_print_meta: f_norm_eps       = 0.0e+00
  llm_load_print_meta: f_norm_rms_eps   = 1.0e-05
  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             = 5632
  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        = 0
  llm_load_print_meta: rope scaling     = linear
  llm_load_print_meta: freq_base_train  = 10000.0
  llm_load_print_meta: freq_scale_train = 1
  llm_load_print_meta: n_ctx_orig_yarn  = 2048
  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      = Q4_0
  llm_load_print_meta: model params     = 1.10 B
  llm_load_print_meta: model size       = 606.53 MiB (4.63 BPW) 
  llm_load_print_meta: general.name     = TinyLlama
  llm_load_print_meta: BOS token        = 1 '<s>'
  llm_load_print_meta: EOS token        = 2 '</s>'
  llm_load_print_meta: UNK token        = 0 '<unk>'
  llm_load_print_meta: PAD token        = 2 '</s>'
  llm_load_print_meta: LF token         = 13 '<0x0A>'
  llm_load_print_meta: max token length = 48
  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 S80, compute capability 2.1, VMM: yes
  llm_load_tensors: ggml ctx size =    0.19 MiB
  llm_load_tensors: offloading 22 repeating layers to GPU
  llm_load_tensors: offloading non-repeating layers to GPU
  llm_load_tensors: offloaded 23/23 layers to GPU
  llm_load_tensors:      MUSA0 buffer size =   571.37 MiB
  llm_load_tensors:        CPU buffer size =    35.16 MiB
  ......................................................................................
  llama_new_context_with_model: n_ctx      = 2048
  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  = 10000.0
  llama_new_context_with_model: freq_scale = 1
  llama_kv_cache_init:      MUSA0 KV buffer size =    44.00 MiB
  llama_new_context_with_model: KV self size  =   44.00 MiB, K (f16):   22.00 MiB, V (f16):   22.00 MiB
  llama_new_context_with_model:  MUSA_Host  output buffer size =     0.12 MiB
  llama_new_context_with_model:      MUSA0 compute buffer size =   148.00 MiB
  llama_new_context_with_model:  MUSA_Host compute buffer size =     8.01 MiB
  llama_new_context_with_model: graph nodes  = 710
  llama_new_context_with_model: graph splits = 2
  llama_init_from_gpt_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
  main: llama threadpool init, n_threads = 6
  

• ./tests/test-backend-ops -> passed full log

    1442/1442 tests passed
    Backend MUSA0: OK
  
  2/2 backends passed
  OK