ClaudeBot/llama.cpp
0
0
Fork 1
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-17 20:31:47 +02:00
Code Issues Packages Projects Releases Wiki Activity

convert: accept the official Puzzle BF16 checkpoint's tensor naming

Browse Source 
The officially distributed BF16 checkpoint (NVIDIA-Nemotron-Labs-3-Puzzle-
75B-A9B-BF16) names the trunk model.* (model.layers.*, model.embeddings,
model.norm_f) where the original release used the NemotronH-style
backbone.*, and spells the router bias e_score_correction_bias instead of
e_score_correction.bias. Normalize both at the top of
NemotronHPuzzleModel.modify_tensors so either checkpoint converts; every
tensor name in the official index (42683 keys, MTP head included) resolves
through the tensor map after normalization.
This commit is contained in:
Yaniss
2026-07-26 11:30:37 +02:00
parent 6174f4268c
commit 189b67fc2c
1 changed files with 10 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
conversion/nemotron.py
+10
View File
@@ -449,6 +449,16 @@ class NemotronHPuzzleModel(NemotronHModel):
self.gguf_writer.add_nextn_predict_layers(len(self.mtp_block_configs))
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# The official BF16 checkpoint (NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16)
# names the trunk "model.*" (model.layers.*, model.embeddings, model.norm_f)
# where the original release used the NemotronH-style "backbone.*", and spells
# the router bias "e_score_correction_bias" instead of "e_score_correction.bias";
# normalize so both convert identically.
if name.startswith("model."):
name = "backbone." + name[len("model."):]
if name.endswith("mixer.gate.e_score_correction_bias"):
name = name[: -len("e_score_correction_bias")] + "e_score_correction.bias"
if name.startswith("mtp."):
# the MTP head has exactly two sub-blocks: bid 0 = attention, bid 1 = moe
assert bid is not None and bid in (0, 1), f"Unexpected MTP tensor: {name}"