LLM-Inference/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-08-12 22:31:11 +04:00
Code Issues Packages Projects Releases Wiki Activity

convert : handle per_layer_config in Gemma4 (transformers 5.15) (#26882)

Browse Source 
* fix: handle nested global_head_dim in Gemma4 config

Gemma-4 E4B models have global_head_dim inside text_config
rather than at the top level. Add fallback to support both layouts.

* fix: add fallback for global_head_dim to support per_layer_config format

* fix: read head_dim only from full_attention layers in per_layer_config and num_global_key_value_heads compatibility

* fix: added fallback for num_global_key_value_heads

* fix: read per_layer_config from root hparams

* fix: delete unused text_config

* cleanup and fixes

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
This commit is contained in:
Wang Zhiyu
2026-08-12 08:05:13 +03:00
committed by GitHub
co-authored by Sigbjørn Skjæret
parent cb27fe9c35
commit 6eff593262
1 changed files with 33 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
conversion/gemma.py
+33 -4
View File
@@ -665,7 +665,18 @@ class Gemma4Model(Gemma3Model):
swa_layers = [t == "sliding_attention" for t in self.hparams["layer_types"]]
self.gguf_writer.add_sliding_window_pattern(swa_layers)
head_dim_full = self.hparams["global_head_dim"]
per_layer_config = self.hparams.get("per_layer_config")
layer_types = self.hparams.get("layer_types", [])
if (head_dim_full := self.hparams.get("global_head_dim")) is None and per_layer_config is not None:
for layer_idx, layer_config in per_layer_config.items():
layer_idx = int(layer_idx)
if layer_idx < len(layer_types):
if layer_types[layer_idx] == "full_attention" and "head_dim" in layer_config:
head_dim_full = layer_config["head_dim"]
break
assert head_dim_full is not None
head_dim_swa = self.hparams["head_dim"]
# correct the head dim for global/swa layers
self.gguf_writer.add_key_length(head_dim_full)
@@ -685,8 +696,14 @@ class Gemma4Model(Gemma3Model):
n_ff_arr = [n_ff if il < first_kv_shared_layer_idx else n_ff * 2 for il in range(self.block_count)]
self.gguf_writer.add_feed_forward_length(n_ff_arr)
# handle num_global_key_value_heads
num_key_value_heads_full = self.hparams.get("num_global_key_value_heads")
if (num_key_value_heads_full := self.hparams.get("num_global_key_value_heads")) is None and per_layer_config is not None:
for layer_idx, layer_config in per_layer_config.items():
layer_idx = int(layer_idx)
if layer_idx < len(layer_types):
if layer_types[layer_idx] == "full_attention" and "num_key_value_heads" in layer_config:
num_key_value_heads_full = layer_config["num_key_value_heads"]
break
num_key_value_heads_swa = self.hparams.get("num_key_value_heads")
if num_key_value_heads_full is not None and num_key_value_heads_swa is not None:
value_arr = [num_key_value_heads_swa if is_swa else num_key_value_heads_full for is_swa in swa_layers]
@@ -708,7 +725,19 @@ class Gemma4Model(Gemma3Model):
# IMPORTANT: this ROPE_FREQS tensor is ONLY used by the full_attention layers
rope_params_full = self.hparams["rope_parameters"]["full_attention"]
assert rope_params_full["rope_type"] == "proportional"
head_dim_full = (self.hparams["global_head_dim"])
per_layer_config = self.hparams.get("per_layer_config")
if (head_dim_full := self.hparams.get("global_head_dim")) is None and per_layer_config is not None:
layer_types = self.hparams.get("layer_types", [])
for layer_idx, layer_config in per_layer_config.items():
layer_idx = int(layer_idx)
if layer_idx < len(layer_types):
if layer_types[layer_idx] == "full_attention" and "head_dim" in layer_config:
head_dim_full = layer_config["head_dim"]
break
assert head_dim_full is not None
partial_rotary_factor_full = rope_params_full["partial_rotary_factor"]
n_rot_full = int(head_dim_full * partial_rotary_factor_full / 2)
n_unrot_full = int(head_dim_full / 2) - n_rot_full