MoE Default use triton's blockwise fp8 in TP Case (#3678)

This commit is contained in:
周周周
2025-08-29 11:07:30 +08:00
committed by GitHub
parent b6edd15d55
commit 17b414c2df
7 changed files with 5 additions and 10 deletions

View File

@@ -65,7 +65,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
# Whether to use fastsafetensor load weight (0 or 1)
"FD_USE_FASTSAFETENSOR": lambda: bool(int(os.getenv("FD_USE_FASTSAFETENSOR", "0"))),
# Whether to use DeepGemm for FP8 blockwise MoE.
"FD_USE_DEEP_GEMM": lambda: bool(int(os.getenv("FD_USE_DEEP_GEMM", "1"))),
"FD_USE_DEEP_GEMM": lambda: bool(int(os.getenv("FD_USE_DEEP_GEMM", "0"))),
# Whether to use aggregate send.
"FD_USE_AGGREGATE_SEND": lambda: bool(int(os.getenv("FD_USE_AGGREGATE_SEND", "0"))),
# Whether to open Trace.