CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : Llama-3_1-Nemotron-Ultra-253B-v1 support (#12843)

Browse Source
This commit is contained in:
ymcki
2025-05-03 23:39:51 +08:00
committed by GitHub
parent 1d36b3670b
commit 3bf785f3ef
3 changed files with 24 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
convert_hf_to_gguf.py
View File

@@ -2123,6 +2123,9 @@ class DeciModel(TextModel):
# if n_heads_in_group is not None, then
# _num_kv_heads[il] is num_attention_head // n_heads_in_group and
# _num_heads[il] is num_attention_head
# ***dummy layer*** for nemotron 253B
# if n_heads_in_group is None and ffn_mult is None
# then _num_kv_heads[il] is 0 and _num_heads[il] is 0 and _ffn_dims is 0
for il in range(len(_block_configs)):
if _block_configs[il]["attention"]["n_heads_in_group"] is None:
if _block_configs[il]["attention"]["replace_with_linear"] is True:
@@ -2134,7 +2137,10 @@ class DeciModel(TextModel):
else:
self._num_kv_heads.append(self.hparams["num_attention_heads"] // _block_configs[il]["attention"]["n_heads_in_group"])
self._num_heads.append(self.hparams["num_attention_heads"])
_ffn_multipliers.append(_block_configs[il]["ffn"]["ffn_mult"])
if _block_configs[il]["ffn"]["ffn_mult"] is None: # dummy layer
_ffn_multipliers.append(0.0)
else:
_ffn_multipliers.append(_block_configs[il]["ffn"]["ffn_mult"])
assert self.block_count == len(self._num_kv_heads)
assert self.block_count == len(self._num_heads)
assert self.block_count == len(_ffn_multipliers)