chore: import upstream snapshot with attribution

2026-07-13 13:34:58 +08:00
commit a203934033
1368 changed files with 175001 additions and 0 deletions
@@ -0,0 +1,9 @@
+# Due to the structural changes made to MoE architecture in `transformers>=5.0`,
+# if you need to apply FP8 quantization to MoE models, please use `megatron export`
+# (compatible with vLLM inference).
+# Reference: https://github.com/modelscope/ms-swift/blob/main/examples/megatron/fp8/quant.sh
+CUDA_VISIBLE_DEVICES=0 \
+swift export \
+    --model Qwen/Qwen2.5-3B-Instruct \
+    --quant_method fp8 \
+    --output_dir Qwen2.5-3B-Instruct-FP8