chore: import upstream snapshot with attribution

examples/megatron/fp8/benchmark.sh

@@ -0,0 +1,39 @@
+# test_env: H20, cuda12.9
+# FP8: 8 * 58GiB 8s/it
+# BF16: 8 * 52GiB 13s/it
+PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
+NPROC_PER_NODE=8 \
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+megatron sft \
+    --model Qwen/Qwen3-14B-FP8 \
+    --save_safetensors true \
+    --fp8_recipe blockwise \
+    --fp8_format e4m3 \
+    --fp8_param_gather true \
+    --dataset 'swift/Chinese-Qwen3-235B-2507-Distill-data-110k-SFT#20000' \
+    --load_from_cache_file true \
+    --tensor_model_parallel_size 4 \
+    --micro_batch_size 1 \
+    --global_batch_size 16 \
+    --packing true \
+    --recompute_granularity selective \
+    --num_train_epochs 1 \
+    --finetune true \
+    --cross_entropy_loss_fusion true \
+    --cross_entropy_fusion_impl native \
+    --lr 1e-5 \
+    --lr_warmup_fraction 0.05 \
+    --min_lr 1e-6 \
+    --output_dir megatron_output/Qwen3-14B-FP8 \
+    --eval_steps 200 \
+    --save_steps 200 \
+    --max_length 8192 \
+    --dataloader_num_workers 8 \
+    --dataset_num_proc 8 \
+    --no_save_optim true \
+    --no_save_rng true \
+    --sequence_parallel true \
+    --use_precision_aware_optimizer true \
+    --exp_avg_dtype bf16 \
+    --exp_avg_sq_dtype bf16 \
+    --attention_backend flash

examples/megatron/fp8/llm.sh

@@ -0,0 +1,60 @@
+# 8 * 95GiB
+# "cuda>=12.9"
+# In this example, FP8 training does not provide any speedup.
+PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
+NPROC_PER_NODE=8 \
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+megatron sft \
+    --model Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 \
+    --save_safetensors true \
+    --fp8_recipe blockwise \
+    --fp8_format e4m3 \
+    --fp8_param_gather true \
+    --dataset 'swift/Chinese-Qwen3-235B-2507-Distill-data-110k-SFT#2000' \
+              'swift/self-cognition#1000' \
+    --load_from_cache_file true \
+    --tensor_model_parallel_size 2 \
+    --expert_model_parallel_size 8 \
+    --moe_permute_fusion true \
+    --moe_grouped_gemm true \
+    --moe_shared_expert_overlap true \
+    --moe_aux_loss_coeff 1e-6 \
+    --micro_batch_size 4 \
+    --global_batch_size 16 \
+    --recompute_granularity full \
+    --recompute_method uniform \
+    --recompute_num_layers 1 \
+    --num_train_epochs 1 \
+    --finetune true \
+    --cross_entropy_loss_fusion true \
+    --lr 1e-5 \
+    --lr_warmup_fraction 0.05 \
+    --min_lr 1e-6 \
+    --output_dir megatron_output/Qwen3-30B-A3B-Instruct-2507-FP8 \
+    --eval_steps 200 \
+    --save_steps 200 \
+    --max_length 2048 \
+    --dataloader_num_workers 8 \
+    --dataset_num_proc 8 \
+    --no_save_optim true \
+    --no_save_rng true \
+    --sequence_parallel true \
+    --moe_expert_capacity_factor 2 \
+    --use_precision_aware_optimizer true \
+    --exp_avg_dtype bf16 \
+    --exp_avg_sq_dtype bf16 \
+    --attention_backend flash \
+    --model_author swift \
+    --model_name swift-robot
+
+# CUDA_VISIBLE_DEVICES=0 \
+# swift infer \
+#     --model megatron_output/Qwen3-30B-A3B-Instruct-2507-FP8/vx-xxx/checkpoint-xxx \
+#     --stream true
+
+# CUDA_VISIBLE_DEVICES=0 \
+# swift infer \
+#     --model megatron_output/Qwen3-30B-A3B-Instruct-2507-FP8/vx-xxx/checkpoint-xxx \
+#     --infer_backend vllm \
+#     --vllm_max_model_len 8192 \
+#     --stream true

examples/megatron/fp8/lora.sh

@@ -0,0 +1,82 @@
+# The generated LoRA delta weights cannot be merged into an FP8 base model via Merge-LoRA.
+# Due to the limited precision of FP8, the LoRA delta will be rounded to 0.
+# However, you can use BF16 weights to perform Merge-LoRA.
+
+# Although the model passed in here is BF16, it will be converted to FP8
+# after being loaded as a Megatron model
+PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
+NPROC_PER_NODE=2 \
+CUDA_VISIBLE_DEVICES=0,1 \
+IMAGE_MAX_TOKEN_NUM=1024 \
+VIDEO_MAX_TOKEN_NUM=128 \
+FPS_MAX_FRAMES=12 \
+megatron sft \
+    --model Qwen/Qwen3.5-4B \
+    --save_safetensors true \
+    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#500' \
+              'AI-ModelScope/alpaca-gpt4-data-en#500' \
+              'swift/self-cognition#500' \
+              'AI-ModelScope/LaTeX_OCR:human_handwrite#2000' \
+    --model_author swift \
+    --model_name swift-robot \
+    --merge_lora false \
+    --linear_decoupled_in_proj true \
+    --load_from_cache_file true \
+    --add_non_thinking_prefix true \
+    --loss_scale ignore_empty_think \
+    --fp8_recipe blockwise \
+    --fp8_format e4m3 \
+    --fp8_param_gather true \
+    --split_dataset_ratio 0.01 \
+    --tuner_type lora \
+    --lora_rank 16 \
+    --lora_alpha 32 \
+    --tensor_model_parallel_size 2 \
+    --micro_batch_size 1 \
+    --global_batch_size 2 \
+    --recompute_granularity full \
+    --recompute_method uniform \
+    --recompute_num_layers 1 \
+    --num_train_epochs 1 \
+    --packing true \
+    --finetune true \
+    --freeze_llm false \
+    --freeze_vit true \
+    --freeze_aligner true \
+    --cross_entropy_loss_fusion true \
+    --lr 1e-4 \
+    --lr_warmup_fraction 0.05 \
+    --min_lr 1e-5 \
+    --output_dir megatron_output/Qwen3.5-4B \
+    --eval_steps 200 \
+    --save_steps 200 \
+    --max_length 4096 \
+    --dataloader_num_workers 8 \
+    --dataset_num_proc 8 \
+    --no_save_optim true \
+    --no_save_rng true \
+    --sequence_parallel true \
+    --mtp_num_layers 1 \
+    --attention_backend flash
+
+# Merge-LoRA
+# FP8 base model + BF16 LoRA inference requires inference framework support
+# Alternatively, you can use BF16 base model + BF16 LoRA for inference
+CUDA_VISIBLE_DEVICES=0 \
+NPROC_PER_NODE=1 \
+megatron export \
+    --model Qwen/Qwen3.5-4B \
+    --adapters megatron_output/Qwen3.5-4B/vx-xxx/checkpoint-xxx \
+    --output_dir megatron_output/Qwen3.5-4B/vx-xxx/checkpoint-xxx-merged \
+    --to_hf true \
+    --linear_decoupled_in_proj true \
+    --mtp_num_layers 1 \
+    --merge_lora true
+
+
+# Inference with merged weights
+CUDA_VISIBLE_DEVICES=0 \
+swift infer \
+    --model megatron_output/Qwen3.5-4B/vx-xxx/checkpoint-xxx-merged \
+    --stream true \
+    --enable_thinking false

examples/megatron/fp8/quant.sh

@@ -0,0 +1,13 @@
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+NPROC_PER_NODE=4 \
+megatron export \
+    --model Qwen/Qwen3.5-35B-A3B \
+    --output_dir Qwen3.5-35B-A3B-FP8 \
+    --to_hf true \
+    --fp8_recipe blockwise \
+    --fp8_format e4m3 \
+    --fp8_param_gather true \
+    --mtp_num_layers 1 \
+    --linear_decoupled_in_proj true \
+    --tensor_model_parallel_size 2 \
+    --pipeline_model_parallel_size 2

examples/megatron/fp8/vlm.sh

@@ -0,0 +1,52 @@
+# 8 * 95GiB
+# In this example, FP8 training does not provide any speedup.
+PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
+OMP_NUM_THREADS=14 \
+NPROC_PER_NODE=8 \
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+IMAGE_MAX_TOKEN_NUM=1024 \
+VIDEO_MAX_TOKEN_NUM=128 \
+FPS_MAX_FRAMES=16 \
+megatron sft \
+    --model Qwen/Qwen3-VL-30B-A3B-Instruct-FP8 \
+    --save_safetensors true \
+    --fp8_recipe blockwise \
+    --fp8_format e4m3 \
+    --fp8_param_gather true \
+    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#10000' \
+              'AI-ModelScope/LaTeX_OCR:human_handwrite#5000' \
+              'swift/VideoChatGPT:Generic#2000' \
+    --load_from_cache_file true \
+    --split_dataset_ratio 0.01 \
+    --moe_permute_fusion true \
+    --tensor_model_parallel_size 2 \
+    --expert_model_parallel_size 8 \
+    --moe_grouped_gemm true \
+    --moe_shared_expert_overlap true \
+    --moe_aux_loss_coeff 1e-6 \
+    --micro_batch_size 1 \
+    --global_batch_size 4 \
+    --recompute_granularity full \
+    --recompute_method uniform \
+    --recompute_num_layers 1 \
+    --num_train_epochs 1 \
+    --finetune true \
+    --cross_entropy_loss_fusion true \
+    --lr 1e-5 \
+    --lr_warmup_fraction 0.05 \
+    --min_lr 1e-6 \
+    --output_dir megatron_output/Qwen3-VL-30B-A3B-Instruct \
+    --eval_steps 500 \
+    --save_steps 500 \
+    --max_length 4096 \
+    --packing true \
+    --dataloader_num_workers 8 \
+    --dataset_num_proc 8 \
+    --no_save_optim true \
+    --no_save_rng true \
+    --sequence_parallel true \
+    --moe_expert_capacity_factor 2 \
+    --use_precision_aware_optimizer true \
+    --exp_avg_dtype bf16 \
+    --exp_avg_sq_dtype bf16 \
+    --attention_backend flash