=========================================================== Multi-node config: NNODES=1 NODE_RANK=0 MASTER_ADDR=127.0.0.1 MASTER_PORT=29573 NPROC (GPUs per node) = 8 GPUS=0,1,2,3,4,5,6,7 Global world size = 8 START_STAGE=1 =========================================================== [config] gradient_checkpointing = DISABLED（40GB A100 + LoRA 推荐关掉加速） [config] audio feature cache = OFF（每个 batch 要 ~400ms 做 mel，强烈建议预计算 cache） =========================================================== [stage1] projector_only (2 epochs, lr=5e-5) → /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage1_projector =========================================================== [05:53:26] Loading train split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy/train.jsonl ... [05:53:45] Loaded 787,277 records. [05:53:45] Loading valid split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy/valid.jsonl ... [05:53:46] Loaded 83,936 records. Dataset train=787,277 valid=83,936 | batch=2 accum=3 | world=8 mode=projector_only [build_model] spatial rates: encoder=10.0 Hz, projector shuffle_factor=4, LLM=2.5 Hz, placeholders per 20s clip=50 [build_model] attn_impl='auto' resolved to 'flash_attention_2' [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [build_model] attn_implementation=flash_attention_2 [05:54:19] Building SpatialBEATs model on CPU ... [05:54:21] SpatialBEATs built. [Stage 1/projector_only] trainable=8 Optimizer groups: projector=8(lr=5.00e-05,wd=0.01) lora=0(lr=5.00e-05) beats=0(lr=5.00e-05) other=0(lr=5.00e-05) Optimizer: steps=32804 warmup=984 lr=5e-05 wd=0.01 =========================================================== Multi-node config: NNODES=1 NODE_RANK=0 MASTER_ADDR=127.0.0.1 MASTER_PORT=29573 NPROC (GPUs per node) = 8 GPUS=0,1,2,3,4,5,6,7 Global world size = 8 START_STAGE=1 =========================================================== [config] gradient_checkpointing = DISABLED（40GB A100 + LoRA 推荐关掉加速） [config] audio feature cache = OFF（每个 batch 要 ~400ms 做 mel，强烈建议预计算 cache） =========================================================== [stage1] projector_only (2 epochs, lr=5e-5) → /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage1_projector =========================================================== =========================================================== Multi-node config: NNODES=1 NODE_RANK=0 MASTER_ADDR=127.0.0.1 MASTER_PORT=29573 NPROC (GPUs per node) = 8 GPUS=0,1,2,3,4,5,6,7 Global world size = 8 START_STAGE=1 =========================================================== [config] gradient_checkpointing = DISABLED（40GB A100 + LoRA 推荐关掉加速） [config] audio feature cache = /apdcephfs_cq12/share_302080740/user/schmittzhu/data/qwen_audio_cache_easy_v2/manifest.json =========================================================== [stage1] projector_only (2 epochs, lr=5e-5) → /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage1_projector =========================================================== Using audio feature cache: /apdcephfs_cq12/share_302080740/user/schmittzhu/data/qwen_audio_cache_easy_v2/manifest.json (entries=398,331, in_memory_max=256) [05:58:14] Loading train split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy/train.jsonl ... [05:58:33] Loaded 787,277 records. [05:58:33] Loading valid split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy/valid.jsonl ... [05:58:34] Loaded 83,936 records. Dataset train=787,277 valid=83,936 | batch=2 accum=3 | world=8 mode=projector_only [build_model] spatial rates: encoder=10.0 Hz, projector shuffle_factor=4, LLM=2.5 Hz, placeholders per 20s clip=50 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [build_model] attn_implementation=sdpa [05:59:19] Building SpatialBEATs model on CPU ... [05:59:21] SpatialBEATs built. [Stage 1/projector_only] trainable=8 Optimizer groups: projector=8(lr=5.00e-05,wd=0.01) lora=0(lr=5.00e-05) beats=0(lr=5.00e-05) other=0(lr=5.00e-05) Optimizer: steps=32804 warmup=984 lr=5e-05 wd=0.01 [step-valid 1000] valid_loss=0.739698 =========================================================== Multi-node config: NNODES=1 NODE_RANK=0 MASTER_ADDR=127.0.0.1 MASTER_PORT=29573 NPROC (GPUs per node) = 8 GPUS=0,1,2,3,4,5,6,7 Global world size = 8 START_STAGE=1 =========================================================== [config] gradient_checkpointing = DISABLED（40GB A100 + LoRA 推荐关掉加速） [config] audio feature cache = /apdcephfs_cq12/share_302080740/user/schmittzhu/data/qwen_audio_cache_easy_v2/manifest.json =========================================================== [stage1] projector_only (2 epochs, lr=5e-5) → /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage1_projector =========================================================== Using audio feature cache: /apdcephfs_cq12/share_302080740/user/schmittzhu/data/qwen_audio_cache_easy_v2/manifest.json (entries=398,331, in_memory_max=256) [07:24:34] Loading train split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy_filtered/train.jsonl ... [07:24:52] Loaded 785,036 records. [07:24:52] Loading valid split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy_filtered/valid.jsonl ... [07:24:54] Loaded 83,220 records. Dataset train=785,036 valid=83,220 | batch=2 accum=3 | world=8 mode=projector_only [build_model] spatial rates: encoder=10.0 Hz, projector shuffle_factor=4, LLM=2.5 Hz, placeholders per 20s clip=50 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0[BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [build_model] attn_implementation=sdpa [07:25:39] Building SpatialBEATs model on CPU ... [07:25:41] SpatialBEATs built. [Stage 1/projector_only] trainable=8 Optimizer groups: projector=8(lr=5.00e-05,wd=0.01) lora=0(lr=5.00e-05) beats=0(lr=5.00e-05) other=0(lr=5.00e-05) Optimizer: steps=32710 warmup=981 lr=5e-05 wd=0.01 [step-valid 1000] valid_loss=0.750919 =========================================================== Multi-node config: NNODES=1 NODE_RANK=0 MASTER_ADDR=127.0.0.1 MASTER_PORT=29573 NPROC (GPUs per node) = 8 GPUS=0,1,2,3,4,5,6,7 Global world size = 8 START_STAGE=1 =========================================================== [config] gradient_checkpointing = DISABLED（40GB A100 + LoRA 推荐关掉加速） [config] audio feature cache = /apdcephfs_cq12/share_302080740/user/schmittzhu/data/qwen_audio_cache_easy_v2/manifest.json =========================================================== [stage1] projector_only (2 epochs, lr=5e-5) → /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage1_projector =========================================================== Using audio feature cache: /apdcephfs_cq12/share_302080740/user/schmittzhu/data/qwen_audio_cache_easy_v2/manifest.json (entries=398,331, in_memory_max=256) [08:10:45] Loading train split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy_filtered/train.jsonl ... [08:11:04] Loaded 785,036 records. [08:11:04] Loading valid split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy_filtered/valid.jsonl ... [08:11:06] Loaded 83,220 records. Dataset train=785,036 valid=83,220 | batch=2 accum=3 | world=8 mode=projector_only [build_model] spatial rates: encoder=10.0 Hz, projector shuffle_factor=4, LLM=2.5 Hz, placeholders per 20s clip=50 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [build_model] attn_implementation=sdpa [08:12:57] Building SpatialBEATs model on CPU ... [08:12:59] SpatialBEATs built. [Stage 1/projector_only] trainable=8 Optimizer groups: projector=8(lr=5.00e-05,wd=0.01) lora=0(lr=5.00e-05) beats=0(lr=5.00e-05) other=0(lr=5.00e-05) Optimizer: steps=32710 warmup=981 lr=5e-05 wd=0.01 [step-valid 1000] valid_loss=0.749426 [step-valid 2000] valid_loss=0.627006 [step-valid 3000] valid_loss=0.589691 [step-valid 4000] valid_loss=0.569729 [step-valid 5000] valid_loss=0.556413 [step-valid 6000] valid_loss=0.547310 [step-valid 7000] valid_loss=0.541301 [step-valid 8000] valid_loss=0.535689 [step-valid 9000] valid_loss=0.532491 [step-valid 10000] valid_loss=0.530133 [step-valid 11000] valid_loss=0.528311 [step-valid 12000] valid_loss=0.526991 [step-valid 13000] valid_loss=0.525880 [step-valid 14000] valid_loss=0.525322 [step-valid 15000] valid_loss=0.524911 [step-valid 16000] valid_loss=0.524455 =========================================================== Multi-node config: NNODES=1 NODE_RANK=0 MASTER_ADDR=127.0.0.1 MASTER_PORT=29573 NPROC (GPUs per node) = 8 GPUS=0,1,2,3,4,5,6,7 Global world size = 8 START_STAGE=1 =========================================================== [config] gradient_checkpointing = DISABLED（40GB A100 + LoRA 推荐关掉加速） [config] audio feature cache = /apdcephfs_cq12/share_302080740/user/schmittzhu/data/qwen_audio_cache_easy_v2/manifest.json =========================================================== [stage1] projector_only (2 epochs, lr=5e-5) → /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage1_projector =========================================================== resume from: /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage1_projector/checkpoints/step_0016000_trainable.pt resume mode: FULL (optimizer + scheduler + step counter restored) Using audio feature cache: /apdcephfs_cq12/share_302080740/user/schmittzhu/data/qwen_audio_cache_easy_v2/manifest.json (entries=398,331, in_memory_max=256) [17:58:15] Loading train split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy_filtered/train.jsonl ... [17:58:33] Loaded 784,953 records. [17:58:33] Loading valid split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy_filtered/valid.jsonl ... [17:58:34] Loaded 83,193 records. Dataset train=784,953 valid=83,193 | batch=2 accum=3 | world=8 mode=projector_only [build_model] spatial rates: encoder=10.0 Hz, projector shuffle_factor=4, LLM=2.5 Hz, placeholders per 20s clip=50 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [build_model] attn_implementation=sdpa [17:59:22] Building SpatialBEATs model on CPU ... [17:59:24] SpatialBEATs built. [Stage 1/projector_only] trainable=8 Optimizer groups: projector=8(lr=5.00e-05,wd=0.01) lora=0(lr=5.00e-05) beats=0(lr=5.00e-05) other=0(lr=5.00e-05) Optimizer: steps=32708 warmup=981 lr=5e-05 wd=0.01 Resumed /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage1_projector/checkpoints/step_0016000_trainable.pt: next_ep=2 gos=0 missing=1802 unexpected=0 [step-valid 1000] valid_loss=0.527116 [step-valid 2000] valid_loss=0.527035 [step-valid 3000] valid_loss=0.526957 [step-valid 4000] valid_loss=0.526922 [step-valid 5000] valid_loss=0.526939 [step-valid 6000] valid_loss=0.526869 [step-valid 7000] valid_loss=0.526825 [step-valid 8000] valid_loss=0.526737 [step-valid 9000] valid_loss=0.526654 [step-valid 10000] valid_loss=0.526668 [step-valid 11000] valid_loss=0.526813 [step-valid 12000] valid_loss=0.526823 [step-valid 13000] valid_loss=0.526684 [step-valid 14000] valid_loss=0.526735 [step-valid 15000] valid_loss=0.526725 [step-valid 16000] valid_loss=0.526771 [epoch 2] train_loss=0.451896 valid_loss=0.522959 valid_em=0.0000 opt_steps=16354 secs=29767.2 -> New best epoch=2 valid_loss=0.522959 Done. best_epoch=2 best_valid_loss=0.522959 output_dir=/apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage1_projector =========================================================== [stage2] encoder_lora (3 epochs, lora_lr=3e-5, proj_lr=1e-5) resume from: /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage1_projector/checkpoints/best_trainable.pt → /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage2_encoder_lora =========================================================== Using audio feature cache: /apdcephfs_cq12/share_302080740/user/schmittzhu/data/qwen_audio_cache_easy_v2/manifest.json (entries=398,331, in_memory_max=256) [02:52:30] Loading train split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy_filtered/train.jsonl ... [02:52:50] Loaded 784,953 records. [02:52:50] Loading valid split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy_filtered/valid.jsonl ... [02:52:51] Loaded 83,193 records. Dataset train=784,953 valid=83,193 | batch=2 accum=3 | world=8 mode=encoder_lora [build_model] spatial rates: encoder=10.0 Hz, projector shuffle_factor=4, LLM=2.5 Hz, placeholders per 20s clip=50 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [build_model] attn_implementation=sdpa [02:53:39] Building SpatialBEATs model on CPU ... [02:53:41] SpatialBEATs built. [Stage 2/encoder_lora] LoRA=112 trainable=232 LoRA (first 20): thinker.model.layers.0.self_attn.k_proj, thinker.model.layers.0.self_attn.o_proj, thinker.model.layers.0.self_attn.q_proj, thinker.model.layers.0.self_attn.v_proj, thinker.model.layers.1.self_attn.k_proj, thinker.model.layers.1.self_attn.o_proj, thinker.model.layers.1.self_attn.q_proj, thinker.model.layers.1.self_attn.v_proj, thinker.model.layers.10.self_attn.k_proj, thinker.model.layers.10.self_attn.o_proj, thinker.model.layers.10.self_attn.q_proj, thinker.model.layers.10.self_attn.v_proj, thinker.model.layers.11.self_attn.k_proj, thinker.model.layers.11.self_attn.o_proj, thinker.model.layers.11.self_attn.q_proj, thinker.model.layers.11.self_attn.v_proj, thinker.model.layers.12.self_attn.k_proj, thinker.model.layers.12.self_attn.o_proj, thinker.model.layers.12.self_attn.q_proj, thinker.model.layers.12.self_attn.v_proj Optimizer groups: projector=8(lr=1.00e-05,wd=0.01) lora=224(lr=3.00e-05) beats=0(lr=3.00e-05) other=0(lr=3.00e-05) Optimizer: steps=49062 warmup=1471 lr=3e-05 wd=0.01 Resumed /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage1_projector/checkpoints/best_trainable.pt: next_ep=1 gos=0 missing=2034 unexpected=8 [step-valid 1000] valid_loss=0.483850 [step-valid 2000] valid_loss=0.411955 [step-valid 3000] valid_loss=0.376576 [step-valid 4000] valid_loss=0.359584 [step-valid 5000] valid_loss=0.347872 [step-valid 6000] valid_loss=0.343363 [step-valid 7000] valid_loss=0.334537 [step-valid 8000] valid_loss=0.330495 [step-valid 9000] valid_loss=0.324756 [step-valid 10000] valid_loss=0.323111 [step-valid 11000] valid_loss=0.320016 [step-valid 12000] valid_loss=0.317173 [step-valid 13000] valid_loss=0.312714 [step-valid 14000] valid_loss=0.311855 [step-valid 15000] valid_loss=0.308691 [step-valid 16000] valid_loss=0.310434 [epoch 1] train_loss=0.328057 valid_loss=0.304027 valid_em=0.1562 opt_steps=16354 secs=31864.8 -> New best epoch=1 valid_loss=0.304027 [step-valid 17000] valid_loss=0.303527 [step-valid 18000] valid_loss=0.303224 [step-valid 19000] valid_loss=0.303543 [step-valid 20000] valid_loss=0.301529 [step-valid 21000] valid_loss=0.299365 [step-valid 22000] valid_loss=0.297152 [step-valid 23000] valid_loss=0.298273 [step-valid 24000] valid_loss=0.299002 [step-valid 25000] valid_loss=0.295180 [step-valid 26000] valid_loss=0.293524 [step-valid 27000] valid_loss=0.292771 [step-valid 28000] valid_loss=0.292354 [step-valid 29000] valid_loss=0.291661 [step-valid 30000] valid_loss=0.290088 [step-valid 31000] valid_loss=0.292185 [step-valid 32000] valid_loss=0.289515 [epoch 2] train_loss=0.247221 valid_loss=0.289178 valid_em=0.1562 opt_steps=16354 secs=32091.9 -> New best epoch=2 valid_loss=0.289178 [step-valid 33000] valid_loss=0.285720 [step-valid 34000] valid_loss=0.285754 [step-valid 35000] valid_loss=0.287495 [step-valid 36000] valid_loss=0.283223 [step-valid 37000] valid_loss=0.283852 [step-valid 38000] valid_loss=0.284259 [step-valid 39000] valid_loss=0.284698 [step-valid 40000] valid_loss=0.283562 [step-valid 41000] valid_loss=0.283898 [step-valid 42000] valid_loss=0.283518 [step-valid 43000] valid_loss=0.282952 [step-valid 44000] valid_loss=0.283511 [step-valid 45000] valid_loss=0.283357 [step-valid 46000] valid_loss=0.282830 [step-valid 47000] valid_loss=0.282979 [step-valid 48000] valid_loss=0.282870 [step-valid 49000] valid_loss=0.282946 [epoch 3] train_loss=0.232710 valid_loss=0.287237 valid_em=0.1562 opt_steps=16354 secs=32650.6 -> New best epoch=3 valid_loss=0.287237 Done. best_epoch=3 best_valid_loss=0.287237 output_dir=/apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage2_encoder_lora =========================================================== [stage3] beats_lora (3 epochs, beats_lr=1e-6, lora_lr=1e-5, proj_lr=5e-6) resume from: /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage2_encoder_lora/checkpoints/best_trainable.pt → /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage3_beats_lora =========================================================== Using audio feature cache: /apdcephfs_cq12/share_302080740/user/schmittzhu/data/qwen_audio_cache_easy_v2/manifest.json (entries=398,331, in_memory_max=256) [07:35:35] Loading train split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy_filtered/train.jsonl ... [07:35:54] Loaded 784,953 records. [07:35:54] Loading valid split: /apdcephfs_cq10/share_1603164/user/schmittzhu/data/process_data/genQA/all_qa_llm_by_difficulty_v2/easy_filtered/valid.jsonl ... [07:35:55] Loaded 83,193 records. Dataset train=784,953 valid=83,193 | batch=2 accum=3 | world=8 mode=beats_lora [build_model] spatial rates: encoder=10.0 Hz, projector shuffle_factor=4, LLM=2.5 Hz, placeholders per 20s clip=50 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [BEATsWrapper] Loaded ckpt: readout_scheme='local_spatial_track' patch_adapter_version='v1' use_trunk_spatial_adapters=False cfg.target_token_rate=10.0 wrapper.encoder_token_rate=10.0 [build_model] attn_implementation=sdpa [07:36:42] Building SpatialBEATs model on CPU ... [07:36:44] SpatialBEATs built. Unfroze spatial_beats_encoder (Stage 3). [Stage 3/beats_lora] LoRA=112 trainable=675 LoRA (first 20): thinker.model.layers.0.self_attn.k_proj, thinker.model.layers.0.self_attn.o_proj, thinker.model.layers.0.self_attn.q_proj, thinker.model.layers.0.self_attn.v_proj, thinker.model.layers.1.self_attn.k_proj, thinker.model.layers.1.self_attn.o_proj, thinker.model.layers.1.self_attn.q_proj, thinker.model.layers.1.self_attn.v_proj, thinker.model.layers.10.self_attn.k_proj, thinker.model.layers.10.self_attn.o_proj, thinker.model.layers.10.self_attn.q_proj, thinker.model.layers.10.self_attn.v_proj, thinker.model.layers.11.self_attn.k_proj, thinker.model.layers.11.self_attn.o_proj, thinker.model.layers.11.self_attn.q_proj, thinker.model.layers.11.self_attn.v_proj, thinker.model.layers.12.self_attn.k_proj, thinker.model.layers.12.self_attn.o_proj, thinker.model.layers.12.self_attn.q_proj, thinker.model.layers.12.self_attn.v_proj Optimizer groups: projector=8(lr=5.00e-06,wd=0.01) lora=224(lr=1.00e-05) beats=443(lr=1.00e-06) other=0(lr=1.00e-05) Optimizer: steps=49062 warmup=1471 lr=1e-05 wd=0.01 Resumed /apdcephfs_cq10/share_1603164/user/schmittzhu/code/Spatial-Qwen/runs/v13d_easy_llmqa/stage2_encoder_lora/checkpoints/best_trainable.pt: next_ep=1 gos=0 missing=1802 unexpected=0