新增支持模型GLM5-MindSpeed-RL-AtomGit

已开启

新增支持模型GLM5 #926

Minds创建于 4月1日从Minds66/MindSpeed-RL： glm5 合入到 Ascend/MindSpeed-RL： master

新增支持模型GLM5 #926

已开启

Minds创建于 4月1日

从Minds66/MindSpeed-RL glm5合入到Ascend/MindSpeed-RL master

共 9 个文件变更+1138-264

Mtests/verl_examples/configs/test_grpo_deepseekv3.2exp_megatron_A3.sh+174-83

	@@ -1,96 +1,187 @@
1	-pkill -9 python	1	+set -x
2	-ray stop --force
3	-rm -rf /tmp/ray
4	-export RAY_DEDUP_LOGS=0
5	-export RAY_DEBUG=0
6	-export RAY_DEBUG_POST_MORTEM=0
7	-export HYDRA_FULL_ERROR=1
8	-export ASCEND_LAUNCH_BLOCKING=0
9	-export ASCEND_RT_VISIBLE_DEVICES='0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15'
10	-export RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES=1
11	-#TASK_QUEUE_ENABLE，下发优化，图模式设置为1，非图模式设置为2
12	-export TASK_QUEUE_ENABLE=1
13	-export HCCL_ASYNC_ERROR_HANDLING=0
14	-export HCCL_EXEC_TIMEOUT=7200
15	-export HCCL_CONNECT_TIMEOUT=7200
16	-export GLOO_CONNECT_TIMEOUT=7200
17	-export HCCL_IF_BASE_PORT=50000
18	-export HCCL_HOST_SOCKET_PORT_RANGE="60000-60050"
19	-export HCCL_NPU_SOCKET_PORT_RANGE="61000-61050"
20	-export HCCL_BUFFSIZE=400
21	-export LD_PRELOAD="/usr/local/lib/libjemalloc.so.2"
22	-export CPU_AFFINITY_CONF=1
23	-export PYTORCH_NPU_ALLOC_CONF="max_split_size_mb:2048"
24	-export PYTHONUNBUFFERED=1
25		2
26	-## VLLM AND CUSTOM	3	+# # 0. download HF checkpoint
27	-export VLLM_VERSION="0.13.0"	4	+# # remove the `quantization_config` in the `config.json`
28	-# 修改为 VLLM_ASCEND 编译后生成的自定义算子路径	5	+# # set `num_nextn_predict_layers=0` to disable MTP, which is not currently supported
29	-export ASCEND_CUSTOM_OPP_PATH='/vllm-ascend/vllm_ascend/_cann_ops_custom/vendors/vllm-ascend'	6	+# huggingface-cli download deepseek-ai/DeepSeek-V3-0324
30	-export LD_LIBRARY_PATH="/vllm-ascend/vllm_ascend/_cann_ops_custom/vendors/vllm-ascend/op_api/lib:$LD_LIBRARY_PATH"	7	+
		8	+# no offline dist checkpoint needed, now with mbridge>=0.13.0, we can directly init model from huggingface downloaded fp8 weights
		9	+# tested on docker://verlai/verl:app-verl0.5-transformers4.55.4-vllm0.10.0-mcore0.13.0-te2.2
		10	+hf_weights="DeepSeek-V3.2-Exp-bf16"
		11	+DIST_CKPT_PATH=""
		12	+
		13	+export RAY_DEDUP_LOGS="0"
31	export VLLM_ASCEND_ENABLE_NZ=0	14	export VLLM_ASCEND_ENABLE_NZ=0
32		15
33	-#修改为当前需要跑的用例路径	16	+# 2. run the script
34	-DEFAULT_SH="./grpo_deepseekv3.2exp_megatron_A3.sh"
35	-echo "Use $DEFAULT_SH"
36		17
37	-ulimit -n 32768	18	+train_files=/dapo-math-17k.parquet
38	-mkdir logs	19	+test_files=/dapo-math-17k.parquet
39		20
40	-export NNODES=16	21	+ALL_OFFLOAD=${ALL_OFFLOAD:-True}
41	-NPUS_PER_NODE=16	22	+COMMON_PARAM_OFFLOAD=${COMMON_PARAM_OFFLOAD:-$ALL_OFFLOAD}
		23	+COMMON_GRAD_OFFLOAD=${COMMON_GRAD_OFFLOAD:-$ALL_OFFLOAD}
		24	+COMMON_OPTIMIZER_OFFLOAD=${COMMON_OPTIMIZER_OFFLOAD:-$ALL_OFFLOAD}
42		25
43	-#修改为当前节点的通信网卡	26	+ACTOR_PARAM_OFFLOAD=${ACTOR_PARAM_OFFLOAD:-$COMMON_PARAM_OFFLOAD}
44	-export SOCKET_IFNAME=""	27	+ACTOR_GRAD_OFFLOAD=${ACTOR_GRAD_OFFLOAD:-$COMMON_GRAD_OFFLOAD}
45	-export HCCL_SOCKET_IFNAME=$SOCKET_IFNAME	28	+ACTOR_OPTIMIZER_OFFLOAD=${ACTOR_OPTIMIZER_OFFLOAD:-$COMMON_OPTIMIZER_OFFLOAD}
46	-export TP_SOCKET_IFNAME=$SOCKET_IFNAME	29	+REF_PARAM_OFFLOAD=${REF_PARAM_OFFLOAD:-$COMMON_PARAM_OFFLOAD}
47	-export GLOO_SOCKET_IFNAME=$SOCKET_IFNAME	30	+CRITIC_PARAM_OFFLOAD=${CRITIC_PARAM_OFFLOAD:-$COMMON_PARAM_OFFLOAD}
48	-	31	+CRITIC_GRAD_OFFLOAD=${CRITIC_GRAD_OFFLOAD:-$COMMON_GRAD_OFFLOAD}
49	-#获取当前节点IP	32	+CRITIC_OPTIMIZER_OFFLOAD=${CRITIC_OPTIMIZER_OFFLOAD:-$COMMON_OPTIMIZER_OFFLOAD}
50	-CURRENT_IP=$(ifconfig $SOCKET_IFNAME \| grep -Eo 'inet (addr:)?([0-9]{1,3}\.){3}[0-9]{1,3}' \| awk '{print $NF}')	33	+RM_PARAM_OFFLOAD=${RM_PARAM_OFFLOAD:-$COMMON_PARAM_OFFLOAD}
51	-#修改为对应主节点IP
52	-MASTER_ADDR=
53	-# export MASTER_ADDR=$CURRENT_IP # 单机
54		34
55		35
56	-if [ "$MASTER_ADDR" = "$CURRENT_IP" ]; then	36	+first_layer=3
57	- # 主节点启动	37	+last_layer=2
58	- ray start --head --port 6379 --dashboard-host=$MASTER_ADDR --node-ip-address=$CURRENT_IP --dashboard-port=8265 --resources='{"NPU": '$NPUS_PER_NODE'}'	38	+# PP=16，[3, 4×14, 2]
59		39
60	- while true; do	40	+NNODES=16
61	- ray_status_output=$(ray status)	41	+PP=16
62	- npu_count=$(echo "$ray_status_output" \| grep -oP '(?<=/)\d+\.\d+(?=\s*NPU)' \| head -n 1)	42	+TP=8
63	- npu_count_int=$(echo "$npu_count" \| awk '{print int($1)}')	43	+EP=16
64	- device_count=$((npu_count_int / $NPUS_PER_NODE))	44	+CP=1
		45	+ETP=1
		46	+INFER_TP=64
		47	+max_num_seqs=128
65		48
66	- # 判断 device_count 是否与 NNODES 相等	49	+experiment_name='dsv3-32nodes'
67	- if [ "$device_count" -eq "$NNODES" ]; then	50	+n_gpus_per_node=16
68	- echo "Ray cluster is ready with $device_count devices (from $npu_count NPU resources), starting Python script."
69	- ray status
70	- bash $DEFAULT_SH
71	- break
72	- else
73	- echo "Waiting for Ray to allocate $NNODES devices. Current device count: $device_count"
74	- sleep 5
75	- fi
76	- done
77	-else
78	- # 子节点尝试往主节点注册ray直到成功
79	- while true; do
80	- # 尝试连接 Ray 集群
81	- ray start --address="$MASTER_ADDR:6379" --resources='{"NPU": '$NPUS_PER_NODE'}' --node-ip-address=$CURRENT_IP
82		51
83	- # 检查连接是否成功
84	- ray status
85	- if [ $? -eq 0 ]; then
86	- echo "Successfully connected to the Ray cluster!"
87	- break
88	- else
89	- echo "Failed to connect to the Ray cluster. Retrying in 5 seconds..."
90	- sleep 5
91	- fi
92	- done
93	-fi
94		52
95	-sleep 600	53	+train_batch_size=128
		54	+ppo_mini_batch_size=64
		55	+n_resp_per_prompt=4
96		56
		57	+balance_batch=False
		58	+
		59	+max_prompt_length=$((1024 * 16))
		60	+max_response_length=$(( 1024 * 8 ))
		61	+total_length=$(($max_prompt_length+$max_response_length))
		62	+
		63	+use_dynamic_bsz=False
		64	+actor_ppo_max_token_len=$(((max_prompt_length + max_response_length) * 1))
		65	+infer_ppo_max_token_len=$(((max_prompt_length + max_response_length) * 1))
		66	+
		67	+use_kl_in_reward=False
		68	+kl_coef=0.0
		69	+use_kl_loss=True
		70	+kl_loss_coef=0.001
		71	+
		72	+clip_ratio_low=0.2
		73	+clip_ratio_high=0.28
		74	+
		75	+exp_name="685B-${NNODES}-train-pp${PP}-tp${TP}-ep${EP}-CP${CP}-actor-length${actor_ppo_max_token_len}_final"
		76	+CKPTS_DIR=${CKPTS_DIR:-"${exp_name}"}
		77	+
		78	+python3 -m verl.trainer.main_ppo \
		79	+ --config-path=./config --config-name='ppo_megatron_trainer'\
		80	+ algorithm.adv_estimator=grpo \
		81	+ data.train_files="$train_files" \
		82	+ data.val_files="$test_files" \
		83	+ data.train_batch_size=$train_batch_size \
		84	+ data.max_prompt_length=$max_prompt_length \
		85	+ data.max_response_length=$max_response_length \
		86	+ data.filter_overlong_prompts=False \
		87	+ data.truncation='left' \
		88	+ actor_rollout_ref.nccl_timeout=7200 \
		89	+ actor_rollout_ref.model.path=$hf_weights \
		90	+ actor_rollout_ref.actor.optim.lr=1e-6 \
		91	+ actor_rollout_ref.actor.ppo_mini_batch_size=${ppo_mini_batch_size} \
		92	+ actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \
		93	+ actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=1 \
		94	+ actor_rollout_ref.actor.use_torch_compile=False \
		95	+ actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
		96	+ actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \
		97	+ actor_rollout_ref.rollout.name=vllm \
		98	+ actor_rollout_ref.rollout.gpu_memory_utilization=0.8 \
		99	+ actor_rollout_ref.rollout.n=${n_resp_per_prompt} \
		100	+ actor_rollout_ref.rollout.temperature=1.0 \
		101	+ actor_rollout_ref.rollout.top_p=1.0 \
		102	+ actor_rollout_ref.rollout.top_k=-1 \
		103	+ actor_rollout_ref.rollout.tensor_model_parallel_size=$INFER_TP \
		104	+ actor_rollout_ref.rollout.load_format='dummy' \
		105	+ actor_rollout_ref.rollout.max_num_seqs=$max_num_seqs \
		106	+ actor_rollout_ref.rollout.calculate_log_probs=True \
		107	+ actor_rollout_ref.actor.clip_ratio_low=${clip_ratio_low} \
		108	+ actor_rollout_ref.actor.clip_ratio_high=${clip_ratio_high} \
		109	+ actor_rollout_ref.actor.clip_ratio_c=10.0 \
		110	+ algorithm.use_kl_in_reward=False \
		111	+ trainer.logger='["console","tensorboard"]' \
		112	+ trainer.project_name='verl_megatron_gsm8k_examples' \
		113	+ trainer.experiment_name=$experiment_name \
		114	+ trainer.n_gpus_per_node=$n_gpus_per_node \
		115	+ trainer.nnodes=$NNODES \
		116	+ trainer.save_freq=-1 \
		117	+ trainer.test_freq=-1 \
		118	+ actor_rollout_ref.model.use_remove_padding=False \
		119	+ actor_rollout_ref.actor.megatron.use_remove_padding=False \
		120	+ actor_rollout_ref.rollout.enforce_eager=False \
		121	+ actor_rollout_ref.model.use_fused_kernels=False \
		122	+ actor_rollout_ref.actor.use_dynamic_bsz=${use_dynamic_bsz} \
		123	+ actor_rollout_ref.ref.log_prob_use_dynamic_bsz=${use_dynamic_bsz} \
		124	+ actor_rollout_ref.rollout.log_prob_use_dynamic_bsz=${use_dynamic_bsz} \
		125	+ actor_rollout_ref.actor.ppo_max_token_len_per_gpu=${actor_ppo_max_token_len} \
		126	+ actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=${infer_ppo_max_token_len} \
		127	+ actor_rollout_ref.rollout.log_prob_max_token_len_per_gpu=${infer_ppo_max_token_len} \
		128	+ actor_rollout_ref.actor.megatron.override_transformer_config.attention_backend='fused' \
		129	+ +actor_rollout_ref.actor.megatron.override_transformer_config.use_flash_attn=True \
		130	+ +actor_rollout_ref.actor.megatron.override_transformer_config.use_distributed_optimizer=True \
		131	+ actor_rollout_ref.actor.strategy=megatron \
		132	+ actor_rollout_ref.actor.megatron.pipeline_model_parallel_size=$PP \
		133	+ actor_rollout_ref.ref.megatron.pipeline_model_parallel_size=$PP \
		134	+ actor_rollout_ref.actor.megatron.tensor_model_parallel_size=$TP \
		135	+ actor_rollout_ref.ref.megatron.tensor_model_parallel_size=$TP \
		136	+ actor_rollout_ref.actor.megatron.expert_model_parallel_size=$EP \
		137	+ actor_rollout_ref.ref.megatron.expert_model_parallel_size=$EP \
		138	+ actor_rollout_ref.actor.megatron.expert_tensor_parallel_size=$ETP \
		139	+ actor_rollout_ref.ref.megatron.expert_tensor_parallel_size=$ETP \
		140	+ actor_rollout_ref.actor.megatron.context_parallel_size=$CP \
		141	+ +actor_rollout_ref.actor.megatron.override_transformer_config.sequence_parallel=True \
		142	+ actor_rollout_ref.actor.megatron.param_offload=True \
		143	+ actor_rollout_ref.actor.megatron.optimizer_offload=True \
		144	+ actor_rollout_ref.actor.megatron.grad_offload=True \
		145	+ actor_rollout_ref.ref.megatron.param_offload=True \
		146	+ +actor_rollout_ref.actor.optim.override_optimizer_config.optimizer_offload_fraction=1 \
		147	+ +actor_rollout_ref.actor.optim.override_optimizer_config.use_precision_aware_optimizer=True \
		148	+ +actor_rollout_ref.actor.optim.override_optimizer_config.optimizer_cpu_offload=True \
		149	+ +actor_rollout_ref.actor.megatron.override_transformer_config.gradient_accumulation_fusion=False \
		150	+ +actor_rollout_ref.actor.megatron.override_transformer_config.masked_softmax_fusion=False \
		151	+ +actor_rollout_ref.actor.megatron.override_transformer_config.bias_dropout_fusion=False \
		152	+ +actor_rollout_ref.actor.megatron.override_transformer_config.num_layers_in_first_pipeline_stage=$first_layer \
		153	+ +actor_rollout_ref.actor.megatron.override_transformer_config.num_layers_in_last_pipeline_stage=$last_layer \
		154	+ +actor_rollout_ref.actor.megatron.override_transformer_config.attention_softmax_in_fp32=True \
		155	+ actor_rollout_ref.rollout.max_num_batched_tokens=$total_length \
		156	+ actor_rollout_ref.actor.megatron.use_mbridge=True \
		157	+ actor_rollout_ref.actor.megatron.vanilla_mbridge=False \
		158	+ actor_rollout_ref.actor.megatron.use_dist_checkpointing=False \
		159	+ actor_rollout_ref.ref.megatron.use_dist_checkpointing=False \
		160	+ actor_rollout_ref.actor.megatron.dist_checkpointing_path=$DIST_CKPT_PATH \
		161	+ actor_rollout_ref.ref.megatron.dist_checkpointing_path=$DIST_CKPT_PATH \
		162	+ trainer.default_local_dir=$CKPTS_DIR \
		163	+ +actor_rollout_ref.actor.megatron.override_transformer_config.recompute_method=uniform \
		164	+ +actor_rollout_ref.actor.megatron.override_transformer_config.recompute_granularity=full \
		165	+ +actor_rollout_ref.actor.megatron.override_transformer_config.recompute_num_layers=1 \
		166	+ +actor_rollout_ref.actor.megatron.override_transformer_config.normalization=RMSNorm \
		167	+ +actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_rmsnorm=True \
		168	+ +actor_rollout_ref.actor.megatron.override_transformer_config.swiglu=True \
		169	+ +actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_swiglu=True \
		170	+ trainer.rollout_data_dir="/rollout_data_dir/$(date +%Y%m%d_%H%M%S)" \
		171	+ +actor_rollout_ref.actor.megatron.override_transformer_config.experimental_attention_variant="dsa" \
		172	+ +actor_rollout_ref.actor.megatron.override_transformer_config.use_dsa_absorb=True \
		173	+ +actor_rollout_ref.actor.megatron.override_transformer_config.dsa_indexer_use_sparse_loss=True \
		174	+ +actor_rollout_ref.actor.megatron.override_transformer_config.dsa_indexer_loss_coeff=0.001 \
		175	+ +actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_lightning_indexer=True \
		176	+ +actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_sparse_flash_attention=True \
		177	+ +actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_lightning_indexer_kl_loss=True \
		178	+ +actor_rollout_ref.actor.megatron.override_transformer_config.moe_router_enable_expert_bias=True \
		179	+ +actor_rollout_ref.actor.megatron.override_transformer_config.context_parallel_size=${CP} \
		180	+ +actor_rollout_ref.actor.megatron.override_transformer_config.context_parallel_algo=kvallgather_cp_algo \
		181	+ +actor_rollout_ref.actor.megatron.override_transformer_config.reset_position_ids=False \
		182	+ +actor_rollout_ref.actor.megatron.override_transformer_config.use_ascend_mc2=False \
		183	+ trainer.resume_mode="disable" \
		184	+ trainer.balance_batch=${balance_batch} \
		185	+ trainer.device=npu \
		186	+ trainer.val_before_train=False \
		187	+ trainer.total_epochs=100 2>&1 \| tee logs/$(date +%Y%m%d_%H%M%S).log

Mtests/verl_examples/grpo/grpo_deepseekv3.2exp_megatron_A3.sh+83-175

	@@ -1,188 +1,96 @@
1	-set -x	1	+pkill -9 python
		2	+ray stop --force
		3	+rm -rf /tmp/ray
		4	+export RAY_DEDUP_LOGS=0
		5	+export RAY_DEBUG=0
		6	+export RAY_DEBUG_POST_MORTEM=0
		7	+export HYDRA_FULL_ERROR=1
		8	+export ASCEND_LAUNCH_BLOCKING=0
		9	+export ASCEND_RT_VISIBLE_DEVICES='0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15'
		10	+export RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES=1
		11	+#TASK_QUEUE_ENABLE，下发优化，图模式设置为1，非图模式设置为2
		12	+export TASK_QUEUE_ENABLE=1
		13	+export HCCL_ASYNC_ERROR_HANDLING=0
		14	+export HCCL_EXEC_TIMEOUT=7200
		15	+export HCCL_CONNECT_TIMEOUT=7200
		16	+export GLOO_CONNECT_TIMEOUT=7200
		17	+export HCCL_IF_BASE_PORT=50000
		18	+export HCCL_HOST_SOCKET_PORT_RANGE="60000-60050"
		19	+export HCCL_NPU_SOCKET_PORT_RANGE="61000-61050"
		20	+export HCCL_BUFFSIZE=400
		21	+export LD_PRELOAD="/usr/local/lib/libjemalloc.so.2"
		22	+export CPU_AFFINITY_CONF=1
		23	+export PYTORCH_NPU_ALLOC_CONF="max_split_size_mb:2048"
		24	+export PYTHONUNBUFFERED=1
2		25
3	-# # 0. download HF checkpoint	26	+## VLLM AND CUSTOM
4	-# # remove the `quantization_config` in the `config.json`	27	+export VLLM_VERSION="0.13.0"
5	-# # set `num_nextn_predict_layers=0` to disable MTP, which is not currently supported	28	+# 修改为 VLLM_ASCEND 编译后生成的自定义算子路径
6	-# huggingface-cli download deepseek-ai/DeepSeek-V3-0324	29	+export ASCEND_CUSTOM_OPP_PATH='/vllm-ascend/vllm_ascend/_cann_ops_custom/vendors/vllm-ascend'
7	-	30	+export LD_LIBRARY_PATH="/vllm-ascend/vllm_ascend/_cann_ops_custom/vendors/vllm-ascend/op_api/lib:$LD_LIBRARY_PATH"
8	-# no offline dist checkpoint needed, now with mbridge>=0.13.0, we can directly init model from huggingface downloaded fp8 weights
9	-# tested on docker://verlai/verl:app-verl0.5-transformers4.55.4-vllm0.10.0-mcore0.13.0-te2.2
10	-hf_weights="DeepSeek-V3.2-Exp-bf16"
11	-DIST_CKPT_PATH=""
12	-
13	-export RAY_DEDUP_LOGS="0"
14	export VLLM_ASCEND_ENABLE_NZ=0	31	export VLLM_ASCEND_ENABLE_NZ=0
15		32
16	-# 2. run the script	33	+#修改为当前需要跑的用例路径
		34	+DEFAULT_SH="./test_grpo_deepseekv3.2exp_megatron_A3.sh"
		35	+echo "Use $DEFAULT_SH"
17		36
18	-train_files=/dapo-math-17k.parquet	37	+ulimit -n 32768
19	-test_files=/dapo-math-17k.parquet	38	+mkdir logs
20		39
21	-ALL_OFFLOAD=${ALL_OFFLOAD:-True}	40	+export NNODES=16
22	-COMMON_PARAM_OFFLOAD=${COMMON_PARAM_OFFLOAD:-$ALL_OFFLOAD}	41	+NPUS_PER_NODE=16
23	-COMMON_GRAD_OFFLOAD=${COMMON_GRAD_OFFLOAD:-$ALL_OFFLOAD}
24	-COMMON_OPTIMIZER_OFFLOAD=${COMMON_OPTIMIZER_OFFLOAD:-$ALL_OFFLOAD}
25		42
26	-ACTOR_PARAM_OFFLOAD=${ACTOR_PARAM_OFFLOAD:-$COMMON_PARAM_OFFLOAD}	43	+#修改为当前节点的通信网卡
27	-ACTOR_GRAD_OFFLOAD=${ACTOR_GRAD_OFFLOAD:-$COMMON_GRAD_OFFLOAD}	44	+export SOCKET_IFNAME=""
28	-ACTOR_OPTIMIZER_OFFLOAD=${ACTOR_OPTIMIZER_OFFLOAD:-$COMMON_OPTIMIZER_OFFLOAD}	45	+export HCCL_SOCKET_IFNAME=$SOCKET_IFNAME
29	-REF_PARAM_OFFLOAD=${REF_PARAM_OFFLOAD:-$COMMON_PARAM_OFFLOAD}	46	+export TP_SOCKET_IFNAME=$SOCKET_IFNAME
30	-CRITIC_PARAM_OFFLOAD=${CRITIC_PARAM_OFFLOAD:-$COMMON_PARAM_OFFLOAD}	47	+export GLOO_SOCKET_IFNAME=$SOCKET_IFNAME
31	-CRITIC_GRAD_OFFLOAD=${CRITIC_GRAD_OFFLOAD:-$COMMON_GRAD_OFFLOAD}	48	+
32	-CRITIC_OPTIMIZER_OFFLOAD=${CRITIC_OPTIMIZER_OFFLOAD:-$COMMON_OPTIMIZER_OFFLOAD}	49	+#获取当前节点IP
33	-RM_PARAM_OFFLOAD=${RM_PARAM_OFFLOAD:-$COMMON_PARAM_OFFLOAD}	50	+CURRENT_IP=$(ifconfig $SOCKET_IFNAME \| grep -Eo 'inet (addr:)?([0-9]{1,3}\.){3}[0-9]{1,3}' \| awk '{print $NF}')
		51	+#修改为对应主节点IP
		52	+MASTER_ADDR=
		53	+# export MASTER_ADDR=$CURRENT_IP # 单机
34		54
35		55
		56	+if [ "$MASTER_ADDR" = "$CURRENT_IP" ]; then
		57	+ # 主节点启动
		58	+ ray start --head --port 6379 --dashboard-host=$MASTER_ADDR --node-ip-address=$CURRENT_IP --dashboard-port=8265 --resources='{"NPU": '$NPUS_PER_NODE'}'
36		59
37	-first_layer=3	60	+ while true; do
38	-last_layer=2	61	+ ray_status_output=$(ray status)
39	-# PP=16，[3, 4×14, 2]	62	+ npu_count=$(echo "$ray_status_output" \| grep -oP '(?<=/)\d+\.\d+(?=\s*NPU)' \| head -n 1)
		63	+ npu_count_int=$(echo "$npu_count" \| awk '{print int($1)}')
		64	+ device_count=$((npu_count_int / $NPUS_PER_NODE))
40		65
41	-NNODES=16	66	+ # 判断 device_count 是否与 NNODES 相等
42	-PP=16	67	+ if [ "$device_count" -eq "$NNODES" ]; then
43	-TP=8	68	+ echo "Ray cluster is ready with $device_count devices (from $npu_count NPU resources), starting Python script."
44	-EP=16	69	+ ray status
45	-CP=1	70	+ bash $DEFAULT_SH
46	-ETP=1	71	+ break
47	-INFER_TP=64	72	+ else
48	-max_num_seqs=128	73	+ echo "Waiting for Ray to allocate $NNODES devices. Current device count: $device_count"
		74	+ sleep 5
		75	+ fi
		76	+ done
		77	+else
		78	+ # 子节点尝试往主节点注册ray直到成功
		79	+ while true; do
		80	+ # 尝试连接 Ray 集群
		81	+ ray start --address="$MASTER_ADDR:6379" --resources='{"NPU": '$NPUS_PER_NODE'}' --node-ip-address=$CURRENT_IP
49		82
50	-experiment_name='dsv3-32nodes'	83	+ # 检查连接是否成功
51	-n_gpus_per_node=16	84	+ ray status
		85	+ if [ $? -eq 0 ]; then
		86	+ echo "Successfully connected to the Ray cluster!"
		87	+ break
		88	+ else
		89	+ echo "Failed to connect to the Ray cluster. Retrying in 5 seconds..."
		90	+ sleep 5
		91	+ fi
		92	+ done
		93	+fi
52		94
		95	+sleep 600
53		96
54	-train_batch_size=128
55	-ppo_mini_batch_size=64
56	-n_resp_per_prompt=4
57	-
58	-balance_batch=False
59	-
60	-max_prompt_length=$((1024 * 16))
61	-max_response_length=$(( 1024 * 8 ))
62	-total_length=$(($max_prompt_length+$max_response_length))
63	-
64	-use_dynamic_bsz=False
65	-actor_ppo_max_token_len=$(((max_prompt_length + max_response_length) * 1))
66	-infer_ppo_max_token_len=$(((max_prompt_length + max_response_length) * 1))
67	-
68	-use_kl_in_reward=False
69	-kl_coef=0.0
70	-use_kl_loss=True
71	-kl_loss_coef=0.001
72	-
73	-clip_ratio_low=0.2
74	-clip_ratio_high=0.28
75	-
76	-exp_name="685B-${NNODES}-train-pp${PP}-tp${TP}-ep${EP}-CP${CP}-actor-length${actor_ppo_max_token_len}_final"
77	-CKPTS_DIR=${CKPTS_DIR:-"${exp_name}"}
78	-
79	-python3 -m verl.trainer.main_ppo \
80	- --config-path=./config --config-name='ppo_megatron_trainer'\
81	- algorithm.adv_estimator=grpo \
82	- data.train_files="$train_files" \
83	- data.val_files="$test_files" \
84	- data.train_batch_size=$train_batch_size \
85	- data.max_prompt_length=$max_prompt_length \
86	- data.max_response_length=$max_response_length \
87	- data.filter_overlong_prompts=False \
88	- data.truncation='left' \
89	- actor_rollout_ref.nccl_timeout=7200 \
90	- actor_rollout_ref.model.path=$hf_weights \
91	- actor_rollout_ref.actor.optim.lr=1e-6 \
92	- actor_rollout_ref.actor.ppo_mini_batch_size=${ppo_mini_batch_size} \
93	- actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \
94	- actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=1 \
95	- actor_rollout_ref.actor.use_torch_compile=False \
96	- actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
97	- actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \
98	- actor_rollout_ref.rollout.name=vllm \
99	- actor_rollout_ref.rollout.gpu_memory_utilization=0.8 \
100	- actor_rollout_ref.rollout.n=${n_resp_per_prompt} \
101	- actor_rollout_ref.rollout.temperature=1.0 \
102	- actor_rollout_ref.rollout.top_p=1.0 \
103	- actor_rollout_ref.rollout.top_k=-1 \
104	- actor_rollout_ref.rollout.tensor_model_parallel_size=$INFER_TP \
105	- actor_rollout_ref.rollout.load_format='dummy' \
106	- actor_rollout_ref.rollout.max_num_seqs=$max_num_seqs \
107	- actor_rollout_ref.rollout.calculate_log_probs=True \
108	- actor_rollout_ref.actor.clip_ratio_low=${clip_ratio_low} \
109	- actor_rollout_ref.actor.clip_ratio_high=${clip_ratio_high} \
110	- actor_rollout_ref.actor.clip_ratio_c=10.0 \
111	- algorithm.use_kl_in_reward=False \
112	- trainer.logger='["console","tensorboard"]' \
113	- trainer.project_name='verl_megatron_gsm8k_examples' \
114	- trainer.experiment_name=$experiment_name \
115	- trainer.n_gpus_per_node=$n_gpus_per_node \
116	- trainer.nnodes=$NNODES \
117	- trainer.save_freq=-1 \
118	- trainer.test_freq=-1 \
119	- actor_rollout_ref.model.use_remove_padding=False \
120	- actor_rollout_ref.actor.megatron.use_remove_padding=False \
121	- actor_rollout_ref.rollout.enforce_eager=False \
122	- actor_rollout_ref.model.use_fused_kernels=False \
123	- actor_rollout_ref.actor.use_dynamic_bsz=${use_dynamic_bsz} \
124	- actor_rollout_ref.ref.log_prob_use_dynamic_bsz=${use_dynamic_bsz} \
125	- actor_rollout_ref.rollout.log_prob_use_dynamic_bsz=${use_dynamic_bsz} \
126	- actor_rollout_ref.actor.ppo_max_token_len_per_gpu=${actor_ppo_max_token_len} \
127	- actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=${infer_ppo_max_token_len} \
128	- actor_rollout_ref.rollout.log_prob_max_token_len_per_gpu=${infer_ppo_max_token_len} \
129	- actor_rollout_ref.actor.megatron.override_transformer_config.attention_backend='fused' \
130	- +actor_rollout_ref.actor.megatron.override_transformer_config.use_flash_attn=True \
131	- +actor_rollout_ref.actor.megatron.override_transformer_config.use_distributed_optimizer=True \
132	- actor_rollout_ref.actor.strategy=megatron \
133	- actor_rollout_ref.actor.megatron.pipeline_model_parallel_size=$PP \
134	- actor_rollout_ref.ref.megatron.pipeline_model_parallel_size=$PP \
135	- actor_rollout_ref.actor.megatron.tensor_model_parallel_size=$TP \
136	- actor_rollout_ref.ref.megatron.tensor_model_parallel_size=$TP \
137	- actor_rollout_ref.actor.megatron.expert_model_parallel_size=$EP \
138	- actor_rollout_ref.ref.megatron.expert_model_parallel_size=$EP \
139	- actor_rollout_ref.actor.megatron.expert_tensor_parallel_size=$ETP \
140	- actor_rollout_ref.ref.megatron.expert_tensor_parallel_size=$ETP \
141	- actor_rollout_ref.actor.megatron.context_parallel_size=$CP \
142	- +actor_rollout_ref.actor.megatron.override_transformer_config.sequence_parallel=True \
143	- actor_rollout_ref.actor.megatron.param_offload=True \
144	- actor_rollout_ref.actor.megatron.optimizer_offload=True \
145	- actor_rollout_ref.actor.megatron.grad_offload=True \
146	- actor_rollout_ref.ref.megatron.param_offload=True \
147	- +actor_rollout_ref.actor.optim.override_optimizer_config.optimizer_offload_fraction=1 \
148	- +actor_rollout_ref.actor.optim.override_optimizer_config.use_precision_aware_optimizer=True \
149	- +actor_rollout_ref.actor.optim.override_optimizer_config.optimizer_cpu_offload=True \
150	- +actor_rollout_ref.actor.megatron.override_transformer_config.gradient_accumulation_fusion=False \
151	- +actor_rollout_ref.actor.megatron.override_transformer_config.masked_softmax_fusion=False \
152	- +actor_rollout_ref.actor.megatron.override_transformer_config.bias_dropout_fusion=False \
153	- +actor_rollout_ref.actor.megatron.override_transformer_config.num_layers_in_first_pipeline_stage=$first_layer \
154	- +actor_rollout_ref.actor.megatron.override_transformer_config.num_layers_in_last_pipeline_stage=$last_layer \
155	- +actor_rollout_ref.actor.megatron.override_transformer_config.attention_softmax_in_fp32=True \
156	- actor_rollout_ref.rollout.max_num_batched_tokens=$total_length \
157	- actor_rollout_ref.actor.megatron.use_mbridge=True \
158	- actor_rollout_ref.actor.megatron.vanilla_mbridge=False \
159	- actor_rollout_ref.actor.megatron.use_dist_checkpointing=False \
160	- actor_rollout_ref.ref.megatron.use_dist_checkpointing=False \
161	- actor_rollout_ref.actor.megatron.dist_checkpointing_path=$DIST_CKPT_PATH \
162	- actor_rollout_ref.ref.megatron.dist_checkpointing_path=$DIST_CKPT_PATH \
163	- trainer.default_local_dir=$CKPTS_DIR \
164	- +actor_rollout_ref.actor.megatron.override_transformer_config.recompute_method=uniform \
165	- +actor_rollout_ref.actor.megatron.override_transformer_config.recompute_granularity=full \
166	- +actor_rollout_ref.actor.megatron.override_transformer_config.recompute_num_layers=1 \
167	- +actor_rollout_ref.actor.megatron.override_transformer_config.normalization=RMSNorm \
168	- +actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_rmsnorm=True \
169	- +actor_rollout_ref.actor.megatron.override_transformer_config.swiglu=True \
170	- +actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_swiglu=True \
171	- trainer.rollout_data_dir="/rollout_data_dir/$(date +%Y%m%d_%H%M%S)" \
172	- +actor_rollout_ref.actor.megatron.override_transformer_config.experimental_attention_variant="dsa" \
173	- +actor_rollout_ref.actor.megatron.override_transformer_config.use_dsa_absorb=True \
174	- +actor_rollout_ref.actor.megatron.override_transformer_config.dsa_indexer_use_sparse_loss=True \
175	- +actor_rollout_ref.actor.megatron.override_transformer_config.dsa_indexer_loss_coeff=0.001 \
176	- +actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_lightning_indexer=True \
177	- +actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_sparse_flash_attention=True \
178	- +actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_lightning_indexer_kl_loss=True \
179	- +actor_rollout_ref.actor.megatron.override_transformer_config.moe_router_enable_expert_bias=True \
180	- +actor_rollout_ref.actor.megatron.override_transformer_config.context_parallel_size=${CP} \
181	- +actor_rollout_ref.actor.megatron.override_transformer_config.context_parallel_algo=kvallgather_cp_algo \
182	- +actor_rollout_ref.actor.megatron.override_transformer_config.reset_position_ids=False \
183	- +actor_rollout_ref.actor.megatron.override_transformer_config.use_ascend_mc2=False \
184	- trainer.resume_mode="disable" \
185	- trainer.balance_batch=${balance_batch} \
186	- trainer.device=npu \
187	- trainer.val_before_train=False \
188	- trainer.total_epochs=100 2>&1 \| tee logs/$(date +%Y%m%d_%H%M%S).log

Atests/verl_examples/grpo/grpo_glm5_megatron_A3.sh+92-0

@@ -0,0 +1,92 @@
	1	+pkill -9 python
	2	+ray stop --force
	3	+rm -rf /tmp/ray
	4	+export RAY_DEDUP_LOGS=0
	5	+export RAY_DEBUG=0
	6	+export RAY_DEBUG_POST_MORTEM=0
	7	+export HYDRA_FULL_ERROR=1
	8	+export ASCEND_LAUNCH_BLOCKING=0
	9	+export ASCEND_RT_VISIBLE_DEVICES='0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15'
	10	+export RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES=1
	11	+#TASK_QUEUE_ENABLE，下发优化，图模式设置为1，非图模式设置为2
	12	+export TASK_QUEUE_ENABLE=1
	13	+export HCCL_ASYNC_ERROR_HANDLING=0
	14	+export HCCL_EXEC_TIMEOUT=7200
	15	+export HCCL_CONNECT_TIMEOUT=7200
	16	+export GLOO_CONNECT_TIMEOUT=7200
	17	+export HCCL_IF_BASE_PORT=50000
	18	+export HCCL_HOST_SOCKET_PORT_RANGE="60000-60050"
	19	+export HCCL_NPU_SOCKET_PORT_RANGE="61000-61050"
	20	+export HCCL_BUFFSIZE=400
	21	+export LD_PRELOAD="/usr/local/lib/libjemalloc.so.2"
	22	+export CPU_AFFINITY_CONF=1
	23	+export PYTORCH_NPU_ALLOC_CONF="max_split_size_mb:2048"
	24	+export PYTHONUNBUFFERED=1
	25	+
	26	+## VLLM AND CUSTOM
	27	+export VLLM_VERSION="0.17.0"
	28	+export VLLM_ASCEND_ENABLE_NZ=0
	29	+
	30	+#修改为当前需要跑的用例路径
	31	+DEFAULT_SH="./test_grpo_glm5_A3.sh"
	32	+echo "Use $DEFAULT_SH"
	33	+ulimit -n 32768
	34	+mkdir logs
	35	+
	36	+export NNODES=32
	37	+NPUS_PER_NODE=16
	38	+
	39	+#修改为当前节点的通信网卡
	40	+export SOCKET_IFNAME=""
	41	+export HCCL_SOCKET_IFNAME=$SOCKET_IFNAME
	42	+export TP_SOCKET_IFNAME=$SOCKET_IFNAME
	43	+export GLOO_SOCKET_IFNAME=$SOCKET_IFNAME
	44	+
	45	+#获取当前节点IP
	46	+CURRENT_IP=$(ifconfig $SOCKET_IFNAME \| grep -Eo 'inet (addr:)?([0-9]{1,3}\.){3}[0-9]{1,3}' \| awk '{print $NF}')
	47	+#修改为对应主节点IP
	48	+MASTER_ADDR=
	49	+# export MASTER_ADDR=$CURRENT_IP # 单机
	50	+
	51	+
	52	+if [ "$MASTER_ADDR" = "$CURRENT_IP" ]; then
	53	+ # 主节点启动
	54	+ ray start --head --port 6379 --dashboard-host=$MASTER_ADDR --node-ip-address=$CURRENT_IP --dashboard-port=8265 --resources='{"NPU": '$NPUS_PER_NODE'}'
	55	+
	56	+ while true; do
	57	+ ray_status_output=$(ray status)
	58	+ npu_count=$(echo "$ray_status_output" \| grep -oP '(?<=/)\d+\.\d+(?=\s*NPU)' \| head -n 1)
	59	+ npu_count_int=$(echo "$npu_count" \| awk '{print int($1)}')
	60	+ device_count=$((npu_count_int / $NPUS_PER_NODE))
	61	+
	62	+ # 判断 device_count 是否与 NNODES 相等
	63	+ if [ "$device_count" -eq "$NNODES" ]; then
	64	+ echo "Ray cluster is ready with $device_count devices (from $npu_count NPU resources), starting Python script."
	65	+ ray status
	66	+ bash $DEFAULT_SH
	67	+ break
	68	+ else
	69	+ echo "Waiting for Ray to allocate $NNODES devices. Current device count: $device_count"
	70	+ sleep 5
	71	+ fi
	72	+ done
	73	+else
	74	+ # 子节点尝试往主节点注册ray直到成功
	75	+ while true; do
	76	+ # 尝试连接 Ray 集群
	77	+ ray start --address="$MASTER_ADDR:6379" --resources='{"NPU": '$NPUS_PER_NODE'}' --node-ip-address=$CURRENT_IP
	78	+
	79	+ # 检查连接是否成功
	80	+ ray status
	81	+ if [ $? -eq 0 ]; then
	82	+ echo "Successfully connected to the Ray cluster!"
	83	+ break
	84	+ else
	85	+ echo "Failed to connect to the Ray cluster. Retrying in 5 seconds..."
	86	+ sleep 5
	87	+ fi
	88	+ done
	89	+fi
	90	+
	91	+sleep 600
	92	+

Averl_npu/docs/model/GLM5.md+125-0

@@ -0,0 +1,125 @@
	1	+# 模型简介
	2	+GLM5是智谱AI与清华大学联合发布的下一代基座模型，定位为从"Vibe Coding"转变成"Agentic Engineering"，精准指出了当下正在发生的范式转移。过去，大模型的主要用法是Vibe Coding：人类写prompt，AI生成代码片段，本质上仍是"辅助工具"。而GLM-5瞄准的是Agentic Engineering——AI自主规划、执行、自我修正，跨越数十个步骤完成端到端的软件工程任务。这不是"写更多代码"，而是"理解工程目标、规划实施路径、处理异常、持续优化"的完整闭环。
	3	+
	4	+GLM-5 全程在华为昇腾芯片上基于 MindSpore 框架训练，实现从美国制程硬件的完全独立，既是技术突破，也是自主可控 AI 基础设施的重要里程碑。
	5	+
	6	+# 环境配套
	7	+
	8	+\| 组件 \| 配套版本 \| 备注 \|
	9	+\| ----------------- \| -------------------- \| ------------------------------------ \|
	10	+\| python \| 3.11 \| \|
	11	+\| pytorch \| 2.9.0 \| \|
	12	+\| vllm \| v0.17.0 \| commit b31e932 \|
	13	+\| vllm-ascend \| v0.17.0rc1 \| commit e20f0b1 \|
	14	+\| verl \| release/v0.7.1 \| commit a35d251 \|
	15	+\| transformers \| v5.4.0 \| commit 276f140 \|
	16	+\| Megatron \| main \| commit 1d462bd \|
	17	+\| Megatron-Bridge \| main \| commit 7cabf71 \|
	18	+\| MindSpeed \| dev分支 \| commit 07056df5 \|
	19	+
	20	+
	21	+# 环境安装
	22	+
	23	+## 安装transformers
	24	+
	25	+```bash
	26	+git clone https://github.com/huggingface/transformers.git -b v5.4.0
	27	+cd transformers
	28	+pip install -e .
	29	+cd ..
	30	+```
	31	+
	32	+## 安装vllm
	33	+
	34	+```bash
	35	+git clone https://github.com/vllm-project/vllm.git -b v0.17.0
	36	+cd vllm
	37	+pip3 install -r requirements/common.txt
	38	+pip3 install -r requirements/build.txt
	39	+pip install torch==2.9.0
	40	+pip install torch_npu==2.9.0
	41	+VLLM_TARGET_DEVICE=empty pip install -v -e .
	42	+cd ..
	43	+```
	44	+
	45	+## 安装vllm_ascend
	46	+
	47	+```bash
	48	+git clone https://github.com/vllm-project/vllm-ascend.git -b v0.17.0rc1
	49	+cd vllm-ascend
	50	+git checkout e20f0b1
	51	+pip install -r requirements-dev.txt
	52	+pip install -v -e .
	53	+cd ..
	54	+```
	55	+
	56	+## 安装verl
	57	+
	58	+```bash
	59	+git clone https://github.com/volcengine/verl.git -b release/v0.7.1
	60	+cd verl
	61	+git checkout a35d251
	62	+pip install -r requirements-npu.txt
	63	+pip install -v -e .
	64	+cd ..
	65	+```
	66	+
	67	+## 安装MindSpeed
	68	+
	69	+```bash
	70	+git clone https://gitcode.com/Ascend/MindSpeed.git -b dev
	71	+cd MindSpeed
	72	+git checkout 07056df5
	73	+pip install -e .
	74	+cd ..
	75	+```
	76	+
	77	+## 安装MindSpeedRL-patch
	78	+
	79	+```bash
	80	+git clone https://gitcode.com/Ascend/MindSpeed-RL.git
	81	+cd vllm-ascend && git apply ../MindSpeed-RL/verl_npu/verl_npu/patch/vllm_ascend/e20f0b1/sfa_v1.patch && cd ..
	82	+cd verl && git apply ../MindSpeed-RL/verl_npu/verl_npu/patch/verl/a35d251/ray_trainer.patch && cd ..
	83	+cd MindSpeed && git apply ../MindSpeed-RL/verl_npu/verl_npu/patch/mindspeed/07056df535/dsa.patch && cd ..
	84	+```
	85	+
	86	+## 安装Megatron-LM
	87	+
	88	+```bash
	89	+git clone https://github.com/NVIDIA/Megatron-LM.git
	90	+cd Megatron-LM
	91	+git checkout 1d462bd37dac21
	92	+git apply ../MindSpeed-RL/verl_npu/verl_npu/patch/megatron-core/1d462bd37dac/glm5.patch
	93	+cp -r megatron ../verl/
	94	+cd ..
	95	+```
	96	+
	97	+## 安装Megatron-Bridge
	98	+
	99	+```bash
	100	+git clone https://github.com/NVIDIA-NeMo/Megatron-Bridge.git
	101	+cd Megatron-Bridge
	102	+git checkout 7cabf71
	103	+git apply ../MindSpeed-RL/verl_npu/verl_npu/patch/megatron-bridge/7cabf71/glm5.patch
	104	+cp -r src/megatron/bridge ../verl/megatron
	105	+cd ..
	106	+```
	107	+
	108	+# 模型运行
	109	+
	110	+```bash
	111	+cd verl
	112	+cp ../MindSpeed-RL/tests/verl_examples/grpo/grpo_glm5_megatron_A3 ./
	113	+```
	114	+
	115	+请参照[verl_npu安装指南](https://gitcode.com/Minds66/MindSpeed-RL/tree/master/verl_npu) 安装内存管理优化库
	116	+ `test_grpo_glm5_megatron_A3.sh`脚本即将到来
	117	+
	118	+* `SOCKET_IFNAME`为当前节点的通信网卡
	119	+* `MASTER_ADDR`为对应主节点IP
	120	+* `NNODES`为使用的节点数
	121	+
	122	+
	123	+```bash
	124	+bash grpo_glm5_megatron_A3.sh
	125	+```

Mverl_npu/docs/model/deepseekv3.2.md+6-6

	@@ -1,4 +1,4 @@
1	-#	1	+# 模型简介
2		2
3	由于DeepSeek-V3.2相关依赖的transformers、Megatron-Bridge适配PR尚未合入，因此需要以 patch 补丁形式将关键适配点单独应用到现有版本中。	3	由于DeepSeek-V3.2相关依赖的transformers、Megatron-Bridge适配PR尚未合入，因此需要以 patch 补丁形式将关键适配点单独应用到现有版本中。
4		4
	@@ -78,7 +78,7 @@ cd ..
78	git clone https://gitcode.com/Ascend/MindSpeed-RL.git	78	git clone https://gitcode.com/Ascend/MindSpeed-RL.git
79	cd transformers && git apply ../MindSpeed-RL/verl_npu/verl_npu/patch/transformers/47b0e478f/transformers.patch && pip install -e . && cd ..	79	cd transformers && git apply ../MindSpeed-RL/verl_npu/verl_npu/patch/transformers/47b0e478f/transformers.patch && pip install -e . && cd ..
80	cd vllm && git apply ../MindSpeed-RL/verl_npu/verl_npu/patch/vllm/72506c98349/common.patch && cd ..	80	cd vllm && git apply ../MindSpeed-RL/verl_npu/verl_npu/patch/vllm/72506c98349/common.patch && cd ..
81	-cd vllm-ascend && git apply --whitespace=fix ../MindSpeed-RL/verl_npu/verl_npu/patch/vllm_ascend/0f812dcc58/sfa_v1.patch && cd..	81	+cd vllm-ascend && git apply --whitespace=fix ../MindSpeed-RL/verl_npu/verl_npu/patch/vllm_ascend/0f812dcc58/sfa_v1.patch && cd ..
82	cd verl && git apply ../MindSpeed-RL/verl_npu/verl_npu/patch/verl/0c06358d6/vllm_async_server.patch && cd ..	82	cd verl && git apply ../MindSpeed-RL/verl_npu/verl_npu/patch/verl/0c06358d6/vllm_async_server.patch && cd ..
83	cd MindSpeed && git apply ../MindSpeed-RL/verl_npu/verl_npu/patch/mindspeed/07056df535/dsa.patch && cd ..	83	cd MindSpeed && git apply ../MindSpeed-RL/verl_npu/verl_npu/patch/mindspeed/07056df535/dsa.patch && cd ..
84	#需手动卸载triton	84	#需手动卸载triton
	@@ -115,16 +115,16 @@ cp ../MindSpeed-RL/tests/verl_examples/configs/test_grpo_deepseekv3.2exp_megatro
115	cp ../MindSpeed-RL/tests/verl_examples/grpo/grpo_deepseekv3.2exp_megatron_A3.sh ./	115	cp ../MindSpeed-RL/tests/verl_examples/grpo/grpo_deepseekv3.2exp_megatron_A3.sh ./
116	```	116	```
117		117
118	-请参照[verl_npu安装指南](https://gitcode.com/Minds66/MindSpeed-RL/tree/master/verl_npu) 安装内存管理优化库	118	+请参照[verl_npu安装指南](https://gitcode.com/Ascend/MindSpeed-RL/tree/master/verl_npu)安装内存管理优化库
119	-修改`test_grpo_deepseekv3.2exp_megatron_A3.sh`脚本中的相应的配置	119	+修改`grpo_deepseekv3.2exp_megatron_A3.sh`脚本中的相应的配置
120		120
121	* `ASCEND_CUSTOM_OPP_PATH`为 VLLM_ASCEND 编译后生成的自定义算子路径	121	* `ASCEND_CUSTOM_OPP_PATH`为 VLLM_ASCEND 编译后生成的自定义算子路径
122	* `SOCKET_IFNAME`为当前节点的通信网卡	122	* `SOCKET_IFNAME`为当前节点的通信网卡
123	* `MASTER_ADDR`为对应主节点IP	123	* `MASTER_ADDR`为对应主节点IP
124	* `NNODES`为使用的节点数	124	* `NNODES`为使用的节点数
125		125
126	-修改`grpo_deepseekv3.2exp_megatron_A3.sh`脚本中的权重`hf_weights`和所用数据集`train_files`和`test_files`	126	+修改`test_grpo_deepseekv3.2exp_megatron_A3.sh`脚本中的权重`hf_weights`和所用数据集`train_files`和`test_files`
127		127
128	```bash	128	```bash
129	-bash test_grpo_deepseekv3.2exp_megatron_A3.sh	129	+bash grpo_deepseekv3.2exp_megatron_A3.sh
130	```	130	```

Averl_npu/verl_npu/patch/megatron-bridge/7cabf71/glm5.patch+560-0

@@ -0,0 +1,560 @@
	1	+diff --git a/src/megatron/bridge/models/__init__.py b/src/megatron/bridge/models/__init__.py
	2	+index 0faff443..b32820e3 100644
	3	+--- a/src/megatron/bridge/models/__init__.py
	4	++++ b/src/megatron/bridge/models/__init__.py
	5	+@@ -64,6 +64,9 @@ from megatron.bridge.models.glm import (
	6	+ GLM45ModelProvider355B,
	7	+ GLMMoEModelProvider,
	8	+ )
	9	++from megatron.bridge.models.glm_moe_dsa import (
	10	++ GLM5Bridge,
	11	++)
	12	+ from megatron.bridge.models.gpt_oss import (
	13	+ GPTOSSBridge,
	14	+ GPTOSSProvider,
	15	+@@ -233,6 +236,7 @@ __all__ = [
	16	+ "GLM45ModelProvider355B",
	17	+ "GLM45AirModelProvider106B",
	18	+ "GLM45Bridge",
	19	++ "GLM5Bridge",
	20	+ "GPTModelProvider",
	21	+ "GPTOSSBridge",
	22	+ "GPTOSSProvider",
	23	+diff --git a/src/megatron/bridge/models/conversion/model_bridge.py b/src/megatron/bridge/models/conversion/model_bridge.py
	24	+index 0f989304..64aa362a 100644
	25	+--- a/src/megatron/bridge/models/conversion/model_bridge.py
	26	++++ b/src/megatron/bridge/models/conversion/model_bridge.py
	27	+@@ -65,6 +65,11 @@ from megatron.bridge.peft.lora import LoRAMerge
	28	+ from megatron.bridge.peft.utils import get_adapter_attributes_from_linear
	29	+ from megatron.bridge.utils.common_utils import print_rank_0
	30	+
	31	++kv_flag = {}
	32	++kv_cnt = 0
	33	++def get_kv_flag():
	34	++ global kv_flag
	35	++ return kv_flag
	36	+
	37	+ logger = logging.getLogger(__name__)
	38	+
	39	+@@ -84,6 +89,17 @@ ADAPTER_NAME_MAP = {
	40	+ }
	41	+
	42	+
	43	++def check_full(param_name):
	44	++ kv_local = get_kv_flag()
	45	++ if "self_attention.linear_k_up_proj.weight" not in param_name and "self_attention.linear_v_up_proj.weight" not in param_name:
	46	++ return True, None
	47	++ name_splits = param_name.split(".")
	48	++ layer_number = str(name_splits[2])
	49	++ if layer_number in kv_flag:
	50	++ return True, layer_number
	51	++ return False, layer_number
	52	++
	53	++
	54	+ class MegatronWeightTuple(NamedTuple):
	55	+ """Tuple representing a Megatron model weight with its metadata."""
	56	+
	57	+@@ -773,7 +789,10 @@ class MegatronModelBridge(Generic[HFPreTrained, ModelProviderTarget, MegatronMod
	58	+ hf_weights = self.maybe_modify_loaded_hf_weight(task.mapping.hf_param, hf_state_dict)
	59	+
	60	+ # 2) Delegate conversion & distribution to the bridge
	61	+- converted_weights = task.mapping.hf_to_megatron(hf_weights, task.megatron_module)
	62	++ if "self_attention.linear_k_up_proj.weight" in task.param_name or "self_attention.linear_v_up_proj.weight" in task.param_name:
	63	++ converted_weights = task.mapping.hf_to_megatron(hf_weights, task.megatron_module, task.param_name)
	64	++ else:
	65	++ converted_weights = task.mapping.hf_to_megatron(hf_weights, task.megatron_module)
	66	+
	67	+ # 3) Copy into Megatron param if this rank received a shard
	68	+ if converted_weights is not None:
	69	+@@ -947,8 +966,44 @@ class MegatronModelBridge(Generic[HFPreTrained, ModelProviderTarget, MegatronMod
	70	+ unwrapped_model = unwrap_model(megatron_model)[0]
	71	+ model_config = unwrapped_model.config
	72	+ embeddings_are_tied = self._share_embeddings_and_output_weights(model_config, unwrapped_model)
	73	++
	74	++ kv_flag_local_2 = get_kv_flag()
	75	+ for task in self._with_progress_tracking(megatron_to_hf_tasks, "Converting to HuggingFace", show_progress):
	76	+ converted_weights_dict = task.mapping.megatron_to_hf(task.param_weight, task.megatron_module)
	77	++ hf_name_tmp = None
	78	++ for hf_name, tensor in converted_weights_dict.items():
	79	++ hf_name_tmp = hf_name
	80	++
	81	++ global kv_cnt
	82	++ flag, layer_num = check_full(task.global_param_name)
	83	++ if not flag:
	84	++ for hf_name, tensor in converted_weights_dict.items():
	85	++ kv_flag_local_2[layer_num] = {}
	86	++ kv_flag_local_2[layer_num][task.global_param_name] = tensor.clone()
	87	++ if torch.distributed.get_rank() == 0:
	88	++ print(f"layer_num{layer_num} , {task.global_param_name}, {kv_cnt} continue")
	89	++ continue
	90	++ elif "self_attention.linear_k_up_proj.weight" in task.global_param_name or "self_attention.linear_v_up_proj.weight" in task.global_param_name:
	91	++ if "self_attention.linear_k_up_proj.weight" in task.global_param_name:
	92	++ for param_name, tensor_v in kv_flag_local_2[layer_num].items():
	93	++ tensor_k = converted_weights_dict[hf_name_tmp].clone()
	94	++ tensor_k = tensor_k.view(64, 192, 512)
	95	++ tensor_v = tensor_v.view(64, 256, 512)
	96	++ new_tensor = torch.cat([tensor_k, tensor_v], dim=1)
	97	++ converted_weights_dict[hf_name_tmp] = new_tensor.view(-1, tensor_k.shape[-1])
	98	++ else:
	99	++ for param_name, tensor_k in kv_flag_local_2[layer_num].items():
	100	++ tensor_v = converted_weights_dict[hf_name_tmp].clone()
	101	++ tensor_k = tensor_k.view(64, 192, 512)
	102	++ tensor_v = tensor_v.view(64, 256, 512)
	103	++ new_tensor = torch.cat([tensor_k, tensor_v], dim=1)
	104	++ converted_weights_dict[hf_name_tmp] = new_tensor.view(-1, tensor_k.shape[-1])
	105	++
	106	++ del kv_flag_local_2[layer_num]
	107	++ if torch.distributed.get_rank() == 0:
	108	++ print(f"layer_num{layer_num} , {task.global_param_name}, {kv_cnt} done")
	109	++ kv_cnt = kv_cnt + 1
	110	++
	111	+ converted_weights_dict = self.maybe_modify_converted_hf_weight(
	112	+ task, converted_weights_dict
	113	+ ) # dict will be none except for one expert;
	114	+diff --git a/src/megatron/bridge/models/conversion/param_mapping.py b/src/megatron/bridge/models/conversion/param_mapping.py
	115	+index 84adb91b..aa63bed2 100644
	116	+--- a/src/megatron/bridge/models/conversion/param_mapping.py
	117	++++ b/src/megatron/bridge/models/conversion/param_mapping.py
	118	+@@ -789,8 +789,21 @@ class ColumnParallelMapping(MegatronParamMapping[torch.Tensor]):
	119	+ self,
	120	+ hf_weights: torch.Tensor,
	121	+ megatron_module: nn.Module,
	122	++ param_name=None
	123	+ ) -> torch.Tensor:
	124	+ """Split weight along dim 0 and distribute to TP ranks."""
	125	++ if param_name is not None and "self_attention.linear_k_up_proj.weight" in param_name:
	126	++ # num_attention_heads, qk_nope_head_dim + v_head_dim, lora_rank(-1)/512
	127	++ hf_weights = hf_weights.reshape(64, 448, -1)
	128	++ # qk_nope_head_dim = 192
	129	++ hf_weights = hf_weights.split([192, 256],dim=1)[0]
	130	++ hf_weights = hf_weights.view(-1, 512)
	131	++ elif param_name is not None and "self_attention.linear_v_up_proj.weight" in param_name:
	132	++ # num_attention_heads, qk_nope_head_dim + v_head_dim, -1
	133	++ hf_weights = hf_weights.reshape(64, 448, -1)
	134	++ # v_head_dim = 256
	135	++ hf_weights = hf_weights.split([192, 256],dim=1)[1]
	136	++ hf_weights = hf_weights.view(-1, 512)
	137	+ if self.tp_size == 1:
	138	+ return hf_weights
	139	+
	140	+@@ -1151,6 +1164,8 @@ class AutoMapping(MegatronParamMapping[torch.Tensor]):
	141	+ def _detect_parallelism_type(self, module: nn.Module) -> str:
	142	+ """Detect parallelism type from module."""
	143	+ module_type = type(module).__name__
	144	++ if module_type.startswith("MindSpeed"):
	145	++ module_type = module_type.replace("MindSpeed", "")
	146	+
	147	+ # Handle fused modules like TELayerNormColumnParallelLinear
	148	+ # These modules have both column-parallel weights (weight, bias)
	149	+@@ -1187,9 +1202,10 @@ class AutoMapping(MegatronParamMapping[torch.Tensor]):
	150	+
	151	+ # Check parallel_mode for TELinear
	152	+ if module_type == "TELinear":
	153	+- if module.parallel_mode == "column":
	154	++ parallel_mode = getattr(module, 'parallel_mode', '')
	155	++ if parallel_mode == "column":
	156	+ return "column"
	157	+- elif module.parallel_mode == "row":
	158	++ elif parallel_mode == "row":
	159	+ return "row"
	160	+ else:
	161	+ return "replicated"
	162	+@@ -1210,6 +1226,7 @@ class AutoMapping(MegatronParamMapping[torch.Tensor]):
	163	+ self,
	164	+ hf_weights: torch.Tensor,
	165	+ megatron_module: nn.Module,
	166	++ param_name=None
	167	+ ) -> torch.Tensor:
	168	+ """Delegate to appropriate mapping based on module type."""
	169	+ # Apply permutation if specified (before distribution)
	170	+@@ -1220,8 +1237,11 @@ class AutoMapping(MegatronParamMapping[torch.Tensor]):
	171	+ if self._mapping is None:
	172	+ self._detected_type = self._detect_parallelism_type(megatron_module)
	173	+ self._mapping = self._get_or_create_mapping(self._detected_type)
	174	+-
	175	+- return self._mapping.hf_to_megatron(hf_weights, megatron_module)
	176	++ if param_name is not None:
	177	++ if "self_attention.linear_k_up_proj.weight" in param_name or "self_attention.linear_v_up_proj.weight" in param_name:
	178	++ return self._mapping.hf_to_megatron(hf_weights, megatron_module, param_name)
	179	++ else:
	180	++ return self._mapping.hf_to_megatron(hf_weights, megatron_module)
	181	+
	182	+ def megatron_to_hf(
	183	+ self,
	184	+diff --git a/src/megatron/bridge/models/glm_moe_dsa/__init__.py b/src/megatron/bridge/models/glm_moe_dsa/__init__.py
	185	+new file mode 100644
	186	+index 00000000..e13c3198
	187	+--- /dev/null
	188	++++ b/src/megatron/bridge/models/glm_moe_dsa/__init__.py
	189	+@@ -0,0 +1,20 @@
	190	++# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
	191	++#
	192	++# Licensed under the Apache License, Version 2.0 (the "License");
	193	++# you may not use this file except in compliance with the License.
	194	++# You may obtain a copy of the License at
	195	++#
	196	++# http://www.apache.org/licenses/LICENSE-2.0
	197	++#
	198	++# Unless required by applicable law or agreed to in writing, software
	199	++# distributed under the License is distributed on an "AS IS" BASIS,
	200	++# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	201	++# See the License for the specific language governing permissions and
	202	++# limitations under the License.
	203	++
	204	++from megatron.bridge.models.glm_moe_dsa.glm5_bridge import GLM5Bridge
	205	++
	206	++
	207	++__all__ = [
	208	++ "GLM5Bridge",
	209	++]
	210	+
	211	+diff --git a/src/megatron/bridge/models/glm_moe_dsa/glm5_bridge.py b/src/megatron/bridge/models/glm_moe_dsa/glm5_bridge.py
	212	+new file mode 100644
	213	+index 00000000..dbe69c34
	214	+--- /dev/null
	215	++++ b/src/megatron/bridge/models/glm_moe_dsa/glm5_bridge.py
	216	+@@ -0,0 +1,195 @@
	217	++import torch
	218	++from megatron.core.models.gpt.gpt_model import GPTModel
	219	++
	220	++from megatron.bridge.models.conversion.mapping_registry import MegatronMappingRegistry
	221	++from megatron.bridge.models.conversion.model_bridge import MegatronModelBridge
	222	++from megatron.bridge.models.conversion.param_mapping import AutoMapping, GatedMLPMapping
	223	++
	224	++from megatron.bridge.models.glm_moe_dsa.glm5_provider import GLM5ModelProvider
	225	++from megatron.bridge.models.hf_pretrained.causal_lm import PreTrainedCausalLM
	226	++
	227	++
	228	++@MegatronModelBridge.register_bridge(source="GlmMoeDsaForCausalLM", target=GPTModel)
	229	++class GLM5Bridge(MegatronModelBridge):
	230	++ """
	231	++ Megatron Bridge for GLM-5 (MoE + MLA + DSA).
	232	++
	233	++ This bridge handles conversion between HuggingFace GlmMoeDsaForCausalLM
	234	++ and Megatron-Core GPTModel formats.
	235	++
	236	++ GLM-5 uses Multi-Latent Attention (MLA), Dynamic Sparse Attention (DSA)
	237	++ indexer layers, and Mixture-of-Experts (MoE).
	238	++ Requires transformers>=5.2.0.
	239	++
	240	++ Example:
	241	++ >>> from megatron.bridge import AutoBridge
	242	++ >>> bridge = AutoBridge.from_hf_pretrained("zai-org/GLM-5")
	243	++ >>> provider = bridge.to_megatron_provider()
	244	++ """
	245	++
	246	++ def provider_bridge(self, hf_pretrained: PreTrainedCausalLM) -> GLM5ModelProvider:
	247	++ hf_config = hf_pretrained.config
	248	++ configs = {
	249	++ "num_layers": hf_config.num_hidden_layers,
	250	++ "hidden_size": hf_config.hidden_size,
	251	++ "ffn_hidden_size": hf_config.intermediate_size,
	252	++ "num_attention_heads": hf_config.num_attention_heads,
	253	++ "num_query_groups": hf_config.num_key_value_heads,
	254	++ "kv_channels": getattr(hf_config, "head_dim", hf_config.hidden_size // hf_config.num_attention_heads),
	255	++ "q_lora_rank": hf_config.q_lora_rank,
	256	++ "kv_lora_rank": hf_config.kv_lora_rank,
	257	++ "num_moe_experts": hf_config.n_routed_experts,
	258	++ "moe_ffn_hidden_size": hf_config.moe_intermediate_size,
	259	++ "moe_shared_expert_intermediate_size": hf_config.moe_intermediate_size * hf_config.n_shared_experts,
	260	++ "moe_layer_freq": [0] * hf_config.first_k_dense_replace + [1] * (hf_config.num_hidden_layers - hf_config.first_k_dense_replace),
	261	++ "moe_router_topk": hf_config.num_experts_per_tok,
	262	++ "moe_router_num_groups": hf_config.n_group,
	263	++ "moe_router_group_topk": hf_config.topk_group,
	264	++ "moe_router_topk_scaling_factor": hf_config.routed_scaling_factor,
	265	++ # MLA dims in MCore format
	266	++ "qk_head_dim": hf_config.qk_nope_head_dim,
	267	++ "qk_pos_emb_head_dim": hf_config.qk_rope_head_dim,
	268	++ "v_head_dim": hf_config.v_head_dim,
	269	++ "vocab_size": hf_config.vocab_size,
	270	++ "rotary_base": hf_config.rope_parameters["rope_theta"],
	271	++ "init_method_std": hf_config.initializer_range,
	272	++ "layernorm_epsilon": hf_config.rms_norm_eps,
	273	++ "multi_latent_attention": True,
	274	++ # DSA indexer params
	275	++ "experimental_attention_variant": "dsa",
	276	++ "dsa_indexer_head_dim": hf_config.index_head_dim,
	277	++ "dsa_indexer_n_heads": hf_config.index_n_heads,
	278	++ "dsa_indexer_topk": hf_config.index_topk,
	279	++ "dsa_indexer_loss_coeff": 0.001,
	280	++ "dsa_indexer_use_sparse_loss": True,
	281	++ # GLM5 uses default rope (no YaRN scaling)
	282	++ "rotary_scaling_factor": 1.0,
	283	++ "mscale": 1.0,
	284	++ "mscale_all_dim": 1.0,
	285	++ }
	286	++
	287	++ dtype = self.dtype_from_hf(hf_config, default=torch.float32)
	288	++ configs["fp16"] = dtype == torch.float16
	289	++ configs["bf16"] = dtype == torch.bfloat16
	290	++ configs["params_dtype"] = dtype
	291	++ configs["make_vocab_size_divisible_by"] = 1280
	292	++ configs["moe_router_score_function"] = "sigmoid"
	293	++ configs["moe_router_enable_expert_bias"] = True
	294	++ if hasattr(hf_config, "aux_loss_alpha"):
	295	++ configs["moe_aux_loss_coeff"] = hf_config.aux_loss_alpha
	296	++
	297	++ provider = GLM5ModelProvider(**configs)
	298	++ # from megatron.core.models.gpt.experimental_attention_variant_module_specs import get_transformer_block_with_experimental_attention_variant_spec)
	299	++ # provider.transformer_layer_spec = get_transformer_block_with_experimental_attention_variant_spec
	300	++
	301	++ provider.normalization = "RMSNorm"
	302	++ provider.gated_linear_unit = True
	303	++ provider.position_embedding_type = "rope"
	304	++ provider.add_bias_linear = False
	305	++ provider.share_embeddings_and_output_weights = False
	306	++ provider.qk_layernorm = True # GLM5 MLA has q_a_layernorm and kv_a_layernorm
	307	++ provider.multi_latent_attention = True
	308	++ provider.moe_grouped_gemm = True
	309	++ provider.moe_router_pre_softmax = True
	310	++ provider.moe_token_dispatcher_type = "alltoall"
	311	++ provider.moe_router_load_balancing_type = "seq_aux_loss"
	312	++ provider.moe_shared_expert_overlap = True
	313	++ provider.moe_router_dtype = "fp32"
	314	++ provider.moe_permute_fusion = True
	315	++ # provider.moe_permute_fusion = False
	316	++ provider.hidden_dropout = 0.0
	317	++ provider.attention_softmax_in_fp32 = False
	318	++
	319	++ return provider
	320	++
	321	++ def mapping_registry(self) -> MegatronMappingRegistry:
	322	++ # mapping_list = get_common_mapping_list() #v32，不用
	323	++ mapping_list = []
	324	++ param_mappings = {
	325	++ # Embed
	326	++ "embedding.word_embeddings.weight": "model.embed_tokens.weight",
	327	++ # LM Head
	328	++ "decoder.final_layernorm.weight": "model.norm.weight",
	329	++ "output_layer.weight": "lm_head.weight",
	330	++ }
	331	++
	332	++ layer_specific_mappings = {
	333	++ # Attention layernorm
	334	++ # "decoder.layers..self_attention.linear_qkv.layer_norm_weight": "model.layers..input_layernorm.weight", # 不用
	335	++ "decoder.layers..input_layernorm.weight": "model.layers..input_layernorm.weight",
	336	++ # Attention output
	337	++ "decoder.layers..self_attention.linear_proj.weight": "model.layers..self_attn.o_proj.weight",
	338	++ # Post-attention layernorm — MoE layers use pre_mlp_layernorm, dense layers use layer_norm_weight
	339	++ "decoder.layers..pre_mlp_layernorm.weight": "model.layers..post_attention_layernorm.weight",
	340	++ "decoder.layers..mlp.linear_fc1.layer_norm_weight": "model.layers..post_attention_layernorm.weight",
	341	++ # MLA weights
	342	++ "decoder.layers..self_attention.linear_q_down_proj.weight": "model.layers..self_attn.q_a_proj.weight",
	343	++ "decoder.layers..self_attention.linear_q_up_proj.weight": "model.layers..self_attn.q_b_proj.weight",
	344	++ "decoder.layers..self_attention.linear_q_up_proj.layer_norm_weight": "model.layers..self_attn.q_a_layernorm.weight",
	345	++ "decoder.layers..self_attention.q_layernorm.weight": "model.layers..self_attn.q_a_layernorm.weight",
	346	++ "decoder.layers..self_attention.linear_kv_down_proj.weight": "model.layers..self_attn.kv_a_proj_with_mqa.weight",
	347	++ # "decoder.layers..self_attention.linear_kv_up_proj.weight": "model.layers..self_attn.kv_b_proj.weight", # DSA
	348	++ "decoder.layers..self_attention.linear_k_up_proj.weight": "model.layers..self_attn.kv_b_proj.weight",# NPU AbsorbedDSA
	349	++ "decoder.layers..self_attention.linear_v_up_proj.weight": "model.layers..self_attn.kv_b_proj.weight",
	350	++ "decoder.layers..self_attention.linear_kv_up_proj.layer_norm_weight": "model.layers..self_attn.kv_a_layernorm.weight",
	351	++ "decoder.layers..self_attention.kv_layernorm.weight": "model.layers..self_attn.kv_a_layernorm.weight",
	352	++ # For non-MLA attention (fallback)
	353	++ "decoder.layers..self_attention.linear_q_proj.weight": "model.layers..self_attn.q_proj.weight",
	354	++ # DSA indexer
	355	++ "decoder.layers..self_attention.core_attention.indexer.linear_wq_b.weight": "model.layers..self_attn.indexer.wq_b.weight",
	356	++ "decoder.layers..self_attention.core_attention.indexer.linear_wk.weight": "model.layers..self_attn.indexer.wk.weight",
	357	++ "decoder.layers..self_attention.core_attention.indexer.k_norm.weight": "model.layers..self_attn.indexer.k_norm.weight",
	358	++ "decoder.layers..self_attention.core_attention.indexer.k_norm.bias": "model.layers..self_attn.indexer.k_norm.bias",
	359	++ "decoder.layers..self_attention.core_attention.indexer.linear_weights_proj.weight": "model.layers..self_attn.indexer.weights_proj.weight",
	360	++ # Dense MLP
	361	++ "decoder.layers..mlp.linear_fc2.weight": "model.layers..mlp.down_proj.weight",
	362	++ # MoE router
	363	++ "decoder.layers..mlp.router.weight": "model.layers..mlp.gate.weight",
	364	++ "decoder.layers..mlp.router.expert_bias": "model.layers..mlp.gate.e_score_correction_bias",
	365	++ # MoE shared experts
	366	++ "decoder.layers..mlp.shared_experts.router.weight": "model.layers..mlp.shared_experts.gate.weight", # 不用
	367	++ "decoder.layers..mlp.shared_experts.linear_fc2.weight": "model.layers..mlp.shared_experts.down_proj.weight",
	368	++ }
	369	++
	370	++ for megatron_param, hf_param in param_mappings.items():
	371	++ mapping_list.append(AutoMapping(megatron_param=megatron_param, hf_param=hf_param))
	372	++
	373	++ for megatron_param, hf_param in layer_specific_mappings.items():
	374	++ mapping_list.append(AutoMapping(megatron_param=megatron_param, hf_param=hf_param))
	375	++
	376	++
	377	++ mapping_list.extend(
	378	++ [
	379	++ # Dense MLP gate+up → fc1
	380	++ GatedMLPMapping(
	381	++ megatron_param="decoder.layers.*.mlp.linear_fc1.weight",
	382	++ gate="model.layers.*.mlp.gate_proj.weight",
	383	++ up="model.layers.*.mlp.up_proj.weight",
	384	++ ),
	385	++
	386	++ # Shared expert gate+up → fc1
	387	++ GatedMLPMapping(
	388	++ megatron_param="decoder.layers.*.mlp.shared_experts.linear_fc1.weight",
	389	++ gate="model.layers.*.mlp.shared_experts.gate_proj.weight",
	390	++ up="model.layers.*.mlp.shared_experts.up_proj.weight",
	391	++ ),
	392	++ ]
	393	++ )
	394	++
	395	++
	396	++ # MoE expert weights (per-expert format: experts.N.gate_proj / up_proj / down_proj)
	397	++ mapping_list.extend(
	398	++ [
	399	++ GatedMLPMapping(
	400	++ megatron_param="decoder.layers..mlp.experts.linear_fc1.weight",
	401	++ gate="model.layers..mlp.experts..gate_proj.weight",
	402	++ up="model.layers..mlp.experts..up_proj.weight",
	403	++ ),
	404	++ # MoE fc2, dsv32这个放上面去了，都一样的。
	405	++ AutoMapping(
	406	++ megatron_param="decoder.layers..mlp.experts.linear_fc2.weight",
	407	++ hf_param="model.layers..mlp.experts..down_proj.weight",
	408	++ ),
	409	++ ]
	410	++ )
	411	++ return MegatronMappingRegistry(*mapping_list)
	412	+
	413	+diff --git a/src/megatron/bridge/models/glm_moe_dsa/glm5_provider.py b/src/megatron/bridge/models/glm_moe_dsa/glm5_provider.py
	414	+new file mode 100644
	415	+index 00000000..cf6e5edc
	416	+--- /dev/null
	417	++++ b/src/megatron/bridge/models/glm_moe_dsa/glm5_provider.py
	418	+@@ -0,0 +1,64 @@
	419	++# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
	420	++#
	421	++# Licensed under the Apache License, Version 2.0 (the "License");
	422	++# you may not use this file except in compliance with the License.
	423	++# You may obtain a copy of the License at
	424	++#
	425	++# http://www.apache.org/licenses/LICENSE-2.0
	426	++#
	427	++# Unless required by applicable law or agreed to in writing, software
	428	++# distributed under the License is distributed on an "AS IS" BASIS,
	429	++# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	430	++# See the License for the specific language governing permissions and
	431	++# limitations under the License.
	432	++
	433	++"""GLM5 uses MLAModelProvider directly. This module is kept for import compatibility."""
	434	++from dataclasses import dataclass
	435	++from functools import partial
	436	++from megatron.bridge.models.gpt_provider import GPTModelProvider
	437	++from megatron.bridge.models.transformer_config import MLATransformerConfig
	438	++
	439	++
	440	++from megatron.core.models.gpt.gpt_layer_specs import get_gpt_decoder_block_spec
	441	++from typing import TYPE_CHECKING, Callable, List, Union
	442	++
	443	++try:
	444	++ import transformer_engine # type: ignore # noqa: F401
	445	++
	446	++ HAVE_TE = True
	447	++except (ImportError, ModuleNotFoundError):
	448	++ HAVE_TE = False
	449	++
	450	++if TYPE_CHECKING:
	451	++ from megatron.core.transformer import ModuleSpec
	452	++
	453	++if HAVE_TE:
	454	++ from megatron.core.utils import is_te_min_version
	455	++
	456	++
	457	++
	458	++@dataclass
	459	++class GLM5ModelProvider(MLATransformerConfig, GPTModelProvider):
	460	++ """Provider for models using Multi-Latent Attention (MLA).
	461	++
	462	++ This class combines MLATransformerConfig (which provides MLA-specific fields
	463	++ like q_lora_rank, kv_lora_rank, qk_head_dim, v_head_dim) with GPTModelProvider
	464	++ (which provides the model instantiation logic).
	465	++
	466	++ Model-specific defaults (normalization, activation, fusions, etc.) should be
	467	++ configured via MEGATRON_DEFAULTS in the respective bridge classes.
	468	++
	469	++ Used by:
	470	++ - DeepSeek V2/V3
	471	++ - Kimi K2
	472	++ - Other MLA-based models
	473	++ """
	474	++ transformer_layer_spec: Union["ModuleSpec", Callable[["GPTModelProvider"], "ModuleSpec"]] = partial(
	475	++ get_gpt_decoder_block_spec, use_transformer_engine=HAVE_TE
	476	++ )
	477	++
	478	++ pass
	479	++
	480	++
	481	++
	482	++__all__ = ["GLM5ModelProvider"]
	483	+
	484	+diff --git a/src/megatron/bridge/models/gpt_provider.py b/src/megatron/bridge/models/gpt_provider.py
	485	+index 6a939800..e5cac258 100644
	486	+--- a/src/megatron/bridge/models/gpt_provider.py
	487	++++ b/src/megatron/bridge/models/gpt_provider.py
	488	+@@ -19,8 +19,6 @@ from dataclasses import dataclass, field
	489	+ from functools import partial
	490	+ from typing import TYPE_CHECKING, Any, Callable, Literal, Optional, Union
	491	+
	492	+-import modelopt.torch.distill as mtd
	493	+-import modelopt.torch.distill.plugins.megatron as mtd_mcore
	494	+ import torch
	495	+ from megatron.core import parallel_state
	496	+ from megatron.core.models.gpt import GPTModel as MCoreGPTModel
	497	+diff --git a/src/megatron/bridge/models/mla_provider.py b/src/megatron/bridge/models/mla_provider.py
	498	+new file mode 100644
	499	+index 00000000..7350d7c5
	500	+--- /dev/null
	501	++++ b/src/megatron/bridge/models/mla_provider.py
	502	+@@ -0,0 +1,44 @@
	503	++# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
	504	++#
	505	++# Licensed under the Apache License, Version 2.0 (the "License");
	506	++# you may not use this file except in compliance with the License.
	507	++# You may obtain a copy of the License at
	508	++#
	509	++# http://www.apache.org/licenses/LICENSE-2.0
	510	++#
	511	++# Unless required by applicable law or agreed to in writing, software
	512	++# distributed under the License is distributed on an "AS IS" BASIS,
	513	++# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	514	++# See the License for the specific language governing permissions and
	515	++# limitations under the License.
	516	++
	517	++"""MLA (Multi-Latent Attention) Model Provider.
	518	++
	519	++This module provides a minimal provider for models using Multi-Latent Attention,
	520	++such as DeepSeek V2/V3 and Kimi K2.
	521	++"""
	522	++
	523	++from dataclasses import dataclass
	524	++
	525	++from megatron.bridge.models.gpt_provider import GPTModelProvider
	526	++from megatron.bridge.models.transformer_config import MLATransformerConfig
	527	++
	528	++
	529	++@dataclass
	530	++class MLAModelProvider(MLATransformerConfig, GPTModelProvider):
	531	++ """Provider for models using Multi-Latent Attention (MLA).
	532	++
	533	++ This class combines MLATransformerConfig (which provides MLA-specific fields
	534	++ like q_lora_rank, kv_lora_rank, qk_head_dim, v_head_dim) with GPTModelProvider
	535	++ (which provides the model instantiation logic).
	536	++
	537	++ Model-specific defaults (normalization, activation, fusions, etc.) should be
	538	++ configured via MEGATRON_DEFAULTS in the respective bridge classes.
	539	++
	540	++ Used by:
	541	++ - DeepSeek V2/V3
	542	++ - Kimi K2
	543	++ - Other MLA-based models
	544	++ """
	545	++
	546	++ pass
	547	+
	548	+diff --git a/src/megatron/bridge/peft/utils.py b/src/megatron/bridge/peft/utils.py
	549	+index 01a67590..ce22314b 100644
	550	+--- a/src/megatron/bridge/peft/utils.py
	551	++++ b/src/megatron/bridge/peft/utils.py
	552	+@@ -60,7 +60,7 @@ HAVE_TE = all(
	553	+ )
	554	+ )
	555	+
	556	+-MixedFusedLayerNorm, HAVE_APEX = safe_import_from("apex.normalization.fused_layer_norm", "MixedFusedLayerNorm")
	557	++HAVE_APEX=False
	558	+
	559	+ TECL = (TEColumnParallelLinear, TELayerNormColumnParallelLinear, TEColumnParallelGroupedLinear)
	560	+ TERL = (TERowParallelLinear, TERowParallelGroupedLinear)

Averl_npu/verl_npu/patch/megatron-core/1d462bd37dac/glm5.patch+72-0

@@ -0,0 +1,72 @@
	1	+diff --git a/megatron/core/models/common/embeddings/rope_utils.py b/megatron/core/models/common/embeddings/rope_utils.py
	2	+index e39540eb1..70783770d 100644
	3	+--- a/megatron/core/models/common/embeddings/rope_utils.py
	4	++++ b/megatron/core/models/common/embeddings/rope_utils.py
	5	+@@ -42,9 +42,30 @@ __all__ = [
	6	+ 'fused_apply_rotary_pos_emb',
	7	+ 'fused_apply_rotary_pos_emb_thd',
	8	+ 'get_pos_emb_on_this_cp_rank',
	9	++ 'apply_rotary_pos_emb_bshd_in_complex',
	10	+ ]
	11	+
	12	+
	13	++def apply_rotary_pos_emb_bshd_in_complex(
	14	++ t: Tensor,
	15	++ freqs: Tensor,
	16	++ rotary_interleaved: bool = False
	17	++ ) -> Tensor:
	18	++ if rotary_interleaved:
	19	++ s, b, n, d = t.shape
	20	++ t = t.view(s, b, n, 2, d // 2).transpose(4, 3)
	21	++ freqs, _ = freqs.chunk(2, dim=-1)
	22	++
	23	++ freqs_cis = torch.polar(torch.ones_like(freqs), freqs)
	24	++
	25	++ x = torch.view_as_complex(t.float().view(*t.shape[:-1], -1, 2)).squeeze(-1)
	26	++ y = torch.view_as_real(x * freqs_cis).flatten(3)
	27	++ if rotary_interleaved:
	28	++ y = torch.cat([y[..., 0::2], y[..., 1::2]], dim=-1)
	29	++
	30	++ return y.to(t.dtype)
	31	++
	32	++
	33	+ def get_pos_emb_on_this_cp_rank(
	34	+ pos_emb: Tensor, seq_dim: int, cp_group: torch.distributed.ProcessGroup
	35	+ ) -> Tensor:
	36	+diff --git a/megatron/core/transformer/experimental_attention_variant/dsa.py b/megatron/core/transformer/experimental_attention_variant/dsa.py
	37	+index 353b31e9b..5cc3e92f4 100644
	38	+--- a/megatron/core/transformer/experimental_attention_variant/dsa.py
	39	++++ b/megatron/core/transformer/experimental_attention_variant/dsa.py
	40	+@@ -13,6 +13,7 @@ from megatron.core.models.common.embeddings import (
	41	+ YarnRotaryEmbedding,
	42	+ apply_rotary_pos_emb,
	43	+ )
	44	++from megatron.core.models.common.embeddings.rope_utils import apply_rotary_pos_emb_bshd_in_complex
	45	+ from megatron.core.packed_seq_params import PackedSeqParams
	46	+ from megatron.core.process_groups_config import ProcessGroupCollection
	47	+ from megatron.core.tensor_parallel.mappings import gather_from_sequence_parallel_region
	48	+@@ -456,19 +457,15 @@ class DSAIndexer(MegatronModule):
	49	+ """Apply RoPE to the input tensor."""
	50	+ # x_nope [seqlen, batch, *, index_head_dim - qk_pos_emb_head_dim]
	51	+ # x_pe [seqlen, batch, *, qk_pos_emb_head_dim]
	52	+- x_nope, x_pe = torch.split(
	53	+- x, [self.index_head_dim - self.qk_pos_emb_head_dim, self.qk_pos_emb_head_dim], dim=-1
	54	++ x_pe, x_nope = torch.split(
	55	++ x, [self.index_head_dim - self.qk_pos_emb_head_dim,self.qk_pos_emb_head_dim], dim=-1
	56	+ )
	57	+- x_pe = apply_rotary_pos_emb(
	58	++ x_pe = apply_rotary_pos_emb_bshd_in_complex(
	59	+ x_pe,
	60	+ rotary_pos_emb,
	61	+- config=self.config,
	62	+- cu_seqlens=None,
	63	+- mscale=mscale,
	64	+- cp_group=self.pg_collection.cp,
	65	+- )
	66	++ rotary_interleaved=True)
	67	+ # [seqlen, batch, *, index_head_dim]
	68	+- x = torch.cat([x_nope, x_pe], dim=-1)
	69	++ x = torch.cat([x_pe, x_nope], dim=-1)
	70	+ return x
	71	+
	72	+ def _compute_index_scores(

Averl_npu/verl_npu/patch/verl/a35d251/ray_trainer.patch+13-0

@@ -0,0 +1,13 @@
	1	+diff --git a/verl/trainer/ppo/ray_trainer.py b/verl/trainer/ppo/ray_trainer.py
	2	+index e178ffc1..fa34f53a 100644
	3	+--- a/verl/trainer/ppo/ray_trainer.py
	4	++++ b/verl/trainer/ppo/ray_trainer.py
	5	+@@ -411,7 +411,7 @@ class RayPPOTrainer:
	6	+
	7	+ lines = []
	8	+ for i in range(n):
	9	+- entry = {k: v[i] for k, v in base_data.items()}
	10	++ entry = {k: str(v[i]) if k == "acc" else v[i] for k, v in base_data.items()}
	11	+ lines.append(json.dumps(entry, ensure_ascii=False))
	12	+
	13	+ with open(filename, "w") as f:

Averl_npu/verl_npu/patch/vllm_ascend/e20f0b1/sfa_v1.patch+13-0

@@ -0,0 +1,13 @@
	1	+diff --git a/vllm_ascend/attention/sfa_v1.py b/vllm_ascend/attention/sfa_v1.py
	2	+index f7edb5fb..33074726 100644
	3	+--- a/vllm_ascend/attention/sfa_v1.py
	4	++++ b/vllm_ascend/attention/sfa_v1.py
	5	+@@ -481,7 +481,7 @@ class AscendSFAImpl(MLAAttentionImpl):
	6	+ # self.W_UV = maybe_trans_nz(self.W_UV)
	7	+
	8	+ # Dispose kv_b_proj since it is replaced by W_UV and W_UK_T to save memory
	9	+- dispose_layer(self.kv_b_proj)
	10	++ #dispose_layer(self.kv_b_proj)
	11	+ if self.enable_dsa_cp:
	12	+ if self.enable_dsa_cp_with_layer_shard:
	13	+ for layer in self.layer_sharding_kwargs or []: