浏览代码

mxfp4 full recipe: switch to the r37 image - r35's DCP4 sparse decode path crashes on this model (reshape x4 DCP factor in attention.py during graph capture)

Alexander Korolev 1 月之前
父节点
当前提交
bbeb6bd553
共有 1 个文件被更改,包括 9 次插入 和 2 次删除
  1. 9 2
      recipes/glm53-full-mxfp4-4x/recipe.toml

+ 9 - 2
recipes/glm53-full-mxfp4-4x/recipe.toml

@@ -137,12 +137,19 @@
 
 engine = "vllm"
 name = "GLM-5.3 (full) MXFP4 (r35 img, RedHatAI ckpt, 4x DCP4)"
-docker_image = "glm53-jovian-r35:local"
+# Image: the r37 DS4.1 build (2026-09-14, night shift) - the r35 tree's DCP4
+# sparse-INDEXER path asserts interleave==1 and then crashes in the DCP sparse
+# decode reshape during graph capture ("shape '[16, 16, 512]' is invalid for
+# input of size 524288" - the x4 DCP factor in the attention.py _sparse_
+# decode). r37 is the newer DS4.1 lock whose DCP sparse-MLA path is the
+# qualified one (same tree family as our DCP2/DFlash2 work); the GlmMoeDsa
+# arch + MTP + b12x ship in the r37 image (verified 2026-09-13).
+docker_image = "jovian-r37-ds41:local"
 docker_image_dir = "glm53-jovian-r27"
 # NOTE: build_profile must stay a TOP-LEVEL key — a bare key after
 # [[extra_mounts]] would parse into the last mount table and be silently
 # ignored (2026-09-08).
-build_profile = "glm53-r35"
+build_profile = "jovian-r37-ds41"
 entrypoint = ["vllm", "serve"]
 model = "RedHatAI/GLM-5.3-MXFP4"
 parent_model = "zai-org/GLM-5.3"