已合并
refactor: modify default qwen3.5 mocre config #2879
refactor: modify default qwen3.5 mocre config #2879
已合并
gaojie创建于 7月24日
共 1 个文件变更+4-32
@@ -7,7 +7,7 @@ gpt_args:
7 context_parallel_algo: ulysses_cp_algo7 context_parallel_algo: ulysses_cp_algo
8 tensor_model_parallel_size: 28 tensor_model_parallel_size: 2
9 sequence_parallel: true9 sequence_parallel: true
10- expert_model_parallel_size: 410+ expert_model_parallel_size: 8
11 pipeline_model_parallel_size: 111 pipeline_model_parallel_size: 1
12 12 
13 ## training:13 ## training:
@@ -91,7 +91,7 @@ model:
91 text_decoder:91 text_decoder:
92 model_id: qwen3_5_moe92 model_id: qwen3_5_moe
93 # parallel93 # parallel
94- expert_model_parallel_size: 494+ expert_model_parallel_size: 8
95 tensor_model_parallel_size: 295 tensor_model_parallel_size: 2
96 pipeline_model_parallel_size: 196 pipeline_model_parallel_size: 1
97 sequence_parallel: true97 sequence_parallel: true
@@ -151,7 +151,7 @@ model:
151 apply_rotary_pos_emb_in_fp32: false151 apply_rotary_pos_emb_in_fp32: false
152 max_position_embeddings: 8192152 max_position_embeddings: 8192
153 vocab_size: 248320153 vocab_size: 248320
154- recompute_num_layers: 0154+ recompute_num_layers: 40
155 recompute_granularity: full155 recompute_granularity: full
156 recompute_method: block156 recompute_method: block
157 patch:157 patch:
@@ -210,35 +210,7 @@ data:
210 dataloader_mode: sampler210 dataloader_mode: sampler
211 drop_last: true211 drop_last: true
212 sampler_type: BaseRandomBatchSampler212 sampler_type: BaseRandomBatchSampler
213- num_workers: 0213+ num_workers: 8
214 collate_param:214 collate_param:
215 model_name: qwen3vl215 model_name: qwen3vl
216 ignore_pad_token_for_loss: true216 ignore_pad_token_for_loss: true
217- 
218- 
219-# 工具配置
220-tools:
221- profile:
222- enable: false
223- profile_type: static
224- ranks: [0]
225- static_param:
226- level: level1
227- with_stack: false
228- with_memory: false
229- record_shapes: false
230- with_cpu: true
231- save_path: ./profiling
232- start_step: 10
233- end_step: 11
234- data_simplification: false
235- aic_metrics_type: PipeUtilization
236- memory_profile:
237- enable: false
238- start_step: 1
239- end_step: 2
240- save_path: ./memory_snapshot
241- dump_ranks: [0]
242- stacks: all
243- max_entries: null
244- mem_info: false