chore: import upstream snapshot with attribution

examples/yaml/deepspeed/infer.sh

@@ -0,0 +1,4 @@
+# Mixed usage
+CUDA_VISIBLE_DEVICES=0 \
+swift infer examples/yaml/deepspeed/infer.yaml \
+    --adapters output/vx-xxx/checkpoint-xxx

examples/yaml/deepspeed/infer.yaml

@@ -0,0 +1,4 @@
+# generation args
+max_new_tokens: 1024
+temperature: 0
+stream: true

examples/yaml/deepspeed/sft.json

@@ -0,0 +1,32 @@
+{
+    "model": "Qwen/Qwen2.5-7B-Instruct",
+    "torch_dtype": "bfloat16",
+    "tuner_type": "lora",
+    "lora_rank": 8,
+    "lora_alpha": 32,
+    "target_modules": "all-linear",
+    "dataset": [
+      "AI-ModelScope/alpaca-gpt4-data-zh#500",
+      "AI-ModelScope/alpaca-gpt4-data-en#500",
+      "swift/self-cognition#500"
+    ],
+    "split_dataset_ratio": 0.0,
+    "max_length": 2048,
+    "system": "You are a helpful assistant.",
+    "model_author": "swift",
+    "model_name": "swift-bot",
+    "num_train_epochs": 1,
+    "per_device_train_batch_size": 1,
+    "per_device_eval_batch_size": 1,
+    "learning_rate": 1e-4,
+    "gradient_accumulation_steps": 8,
+    "eval_steps": 50,
+    "save_steps": 50,
+    "save_total_limit": 2,
+    "logging_steps": 5,
+    "output_dir": "output",
+    "warmup_ratio": 0.05,
+    "dataloader_num_workers": 4,
+    "dataset_num_proc": 4,
+    "deepspeed": "zero2"
+}

examples/yaml/deepspeed/sft.sh

@@ -0,0 +1,7 @@
+NPROC_PER_NODE=2 \
+CUDA_VISIBLE_DEVICES=0,1 \
+swift sft examples/yaml/deepspeed/sft.yaml
+
+# NPROC_PER_NODE=2 \
+# CUDA_VISIBLE_DEVICES=0,1 \
+# swift sft examples/yaml/deepspeed/sft.json

examples/yaml/deepspeed/sft.yaml

@@ -0,0 +1,36 @@
+
+# model args
+model: Qwen/Qwen2.5-7B-Instruct
+torch_dtype: bfloat16
+
+tuner_type: lora
+lora_rank: 8
+lora_alpha: 32
+target_modules: all-linear
+
+# dataset args
+dataset:
+  - 'AI-ModelScope/alpaca-gpt4-data-zh#500'
+  - 'AI-ModelScope/alpaca-gpt4-data-en#500'
+  - 'swift/self-cognition#500'
+split_dataset_ratio: 0.0
+max_length: 2048
+system: 'You are a helpful assistant.'
+model_author: swift
+model_name: swift-bot
+
+# training args
+num_train_epochs: 1
+per_device_train_batch_size: 1
+per_device_eval_batch_size: 1
+learning_rate: 1e-4
+gradient_accumulation_steps: 8
+eval_steps: 50
+save_steps: 50
+save_total_limit: 2
+logging_steps: 5
+output_dir: output
+warmup_ratio: 0.05
+dataloader_num_workers: 4
+dataset_num_proc: 4
+deepspeed: zero2

examples/yaml/megatron/infer.sh

@@ -0,0 +1,4 @@
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+swift infer \
+    examples/yaml/megatron/infer.yaml \
+    --model megatron_output/Qwen3.5-35B-A3B/vx-xxx/checkpoint-xxx-merged \

examples/yaml/megatron/infer.yaml

@@ -0,0 +1,10 @@
+ENV:
+  PYTORCH_CUDA_ALLOC_CONF: 'expandable_segments:True'
+  MAX_PIXELS: '1003520'
+  VIDEO_MAX_PIXELS: '50176'
+  FPS_MAX_FRAMES: '12'
+
+stream: true
+enable_thinking: false
+max_new_tokens: 512
+load_data_args: true

examples/yaml/megatron/sft.sh

@@ -0,0 +1,4 @@
+NPROC_PER_NODE=4 \
+CUDA_VISIBLE_DEVICES=0,1,2,3 \
+megatron sft \
+    examples/yaml/megatron/sft.yaml

examples/yaml/megatron/sft.yaml

@@ -0,0 +1,66 @@
+ENV:
+  PYTORCH_CUDA_ALLOC_CONF: 'expandable_segments:True'
+  MAX_PIXELS: '1003520'
+  VIDEO_MAX_PIXELS: '50176'
+  FPS_MAX_FRAMES: '12'
+
+# model args
+model: "Qwen/Qwen3.5-35B-A3B"
+
+tuner_type: lora
+lora_rank: 8
+lora_alpha: 32
+target_modules: all-linear
+
+# dataset args
+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'
+load_from_cache_file: true
+split_dataset_ratio: 0.01
+max_length: 2048
+dataloader_num_workers: 4
+dataset_num_proc: 4
+model_author: swift
+model_name: swift-bot
+padding_free: true
+packing: true
+
+# template args
+loss_scale: ignore_empty_think
+add_non_thinking_prefix: true
+
+# training args
+save_safetensors: true
+merge_lora: true
+micro_batch_size: 1
+global_batch_size: 4
+num_train_epochs: 1
+finetune: true
+lr: 1e-4
+lr_warmup_fraction: 0.05
+min_lr: 1e-5
+output_dir: megatron_output/Qwen3.5-35B-A3B
+eval_steps: 200
+save_steps: 200
+no_save_optim: true
+no_save_rng: true
+
+freeze_llm: false
+freeze_vit: true
+freeze_aligner: true
+
+expert_model_parallel_size: 4
+sequence_parallel: true
+moe_permute_fusion: true
+moe_grouped_gemm: true
+moe_shared_expert_overlap: true
+moe_aux_loss_coeff: 1e-6
+
+recompute_granularity: full
+recompute_method: uniform
+recompute_num_layers: 1
+cross_entropy_loss_fusion: true
+attention_backend: flash