Feature DeepSeek V3/R1 INT8 Quantization (block-wise) (#3730)

Co-authored-by: HandH1998 <1335248067@qq.com>
2025-02-24 21:43:35 +08:00
parent b110084654
commit 1a6e97577a
9 changed files with 1092 additions and 5 deletions
--- a/python/sglang/srt/layers/linear.py
+++ b/python/sglang/srt/layers/linear.py
@@ -38,6 +38,7 @@ WEIGHT_LOADER_V2_SUPPORTED = [
    "AWQLinearMethod",
    "GPTQMarlinLinearMethod",
    "Fp8LinearMethod",
+    "BlockInt8LinearMethod",
    "MarlinLinearMethod",
    "QQQLinearMethod",
    "GPTQMarlin24LinearMethod",