Cherry pick weight_checker _weight_fp32 buffer skip from #22663 (#24534)

Co-authored-by: JD <jaedon.guo@gmail.com>
2026-05-06 21:21:12 +08:00
parent 0d40931b08
commit 491051c622
1 changed files with 2 additions and 1 deletions
@@ -38,7 +38,7 @@ class WeightChecker:

    def _reset_tensors(self):
        for name, param in self._model_state():
-            if "cos_sin_cache" in name or "freqs_cis" in name:
+            if "cos_sin_cache" in name or "freqs_cis" in name or "_weight_fp32" in name:
                continue
            param.copy_(_random_like(param))

@@ -131,6 +131,7 @@ def _postprocess_tensors(
    non_persistent_buffer_patterns = [
        "cos_sin_cache",  # RoPE cache
        "inv_freq",  # RoPE inverse frequency (if it exists as buffer)
+        "_weight_fp32",  # FP32 cache of gate weight (e.g. Glm4MoeGate)
    ]

    for name in raw: