| 123456789101112131415161718192021222324252627282930313233343536373839 |
- --- a/vllm/modelopt.py
- +++ b/vllm/modelopt.py
- @@ -2356,6 +2356,22 @@
- def _quantized_layer_prefix_candidates(prefix: str) -> tuple[str, ...]:
- candidates = [prefix]
-
- + # MTP draft naming (see nvfp4 stand-up notes): strip mtp_block,
- + # then re-root over every naming the shared quantized_layers dict
- + # can be in - checkpoint form (model.language_model.*) or the
- + # target-tree form its mapper rewrites to (language_model.model.*).
- + _bases = {prefix, prefix.replace(".mtp_block.", ".")}
- + for _b in list(_bases):
- + for _root in ("model.", "model.language_model.", "language_model.model."):
- + if _b.startswith(_root):
- + _bases.add(_b[len(_root) :])
- + for _b in list(_bases):
- + if _b.startswith("layers."):
- + _bases.add("model." + _b)
- + _bases.add("model.language_model." + _b)
- + _bases.add("language_model.model." + _b)
- + candidates.extend(_bases - {prefix})
- +
- if prefix.endswith(".lm_head"):
- candidates.append("lm_head")
-
- @@ -2387,6 +2403,13 @@
- return None
-
- quant_algo = self._resolve_quant_algo(prefix)
- + import sys as _sys
- + print(
- + "[quantprobe] layer=%s prefix=%s algo=%s"
- + % (type(layer).__name__, prefix, quant_algo),
- + file=_sys.stderr,
- + flush=True,
- + )
-
- if isinstance(layer, (LinearBase, ParallelLMHead)):
- if quant_algo == "FP8":
|
