biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Feature][OCP MX] Support mxfp6 and mixed mxfp6-mxfp4 (#21166)

Browse Source
This commit is contained in:
fxmarty-amd
2025-10-07 15:35:26 +02:00
committed by GitHub
parent 08d26a1b7e
commit 41f1cf38f2
18 changed files with 656 additions and 180 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
vllm/scalar_type.py
View File

@@ -337,8 +337,11 @@ class scalar_types:
float16_e5m10 = ScalarType.float_IEEE754(5, 10)
# fp6, https://github.com/usyd-fsalab/fp6_llm/tree/main
# and https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf
float6_e3m2f = ScalarType.float_(3, 2, True, NanRepr.NONE)
float6_e2m3f = ScalarType.float_(2, 3, True, NanRepr.NONE)
# fp4, https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf
float4_e2m1f = ScalarType.float_(2, 1, True, NanRepr.NONE)