nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 24 Packages Projects Releases Wiki Activity

convert : fix Gemma3N, GraniteMoe and Ernie4.5Moe (#19084)

Browse Source 
* fix Gemma3N and Ernie4.5Moe

* fix GraniteMoe
This commit is contained in:
Sigbjørn Skjæret
2026-01-25 13:05:05 +01:00
committed by GitHub
parent 080b161995
commit 70d860824a
1 changed files with 12 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
convert_hf_to_gguf.py
+12 -12
View File
@@ -3799,7 +3799,7 @@ class Ernie4_5MoeModel(Ernie4_5Model):
merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight" merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
yield from super().modify_tensors(data_torch, merged_name, bid) yield from super().modify_tensors(data_torch, merged_name, bid)
else: else:
yield from super().modify_tensors(data_torch, name, bid) yield from ModelBase.modify_tensors(self, data_torch, name, bid)
def prepare_tensors(self): def prepare_tensors(self):
super().prepare_tensors() super().prepare_tensors()
@@ -6153,7 +6153,7 @@ class Gemma3nVisionAudioModel(ConformerAudioModel):
if new_name.endswith("conv_stem.conv.bias") or new_name.endswith("layer_scale.gamma"): if new_name.endswith("conv_stem.conv.bias") or new_name.endswith("layer_scale.gamma"):
data_torch = data_torch.unsqueeze(0).unsqueeze(-1).unsqueeze(-1) # [1, C, 1, 1] data_torch = data_torch.unsqueeze(0).unsqueeze(-1).unsqueeze(-1) # [1, C, 1, 1]
yield from super().modify_tensors(data_torch, new_name, bid) yield from ModelBase.modify_tensors(self, data_torch, new_name, bid)
@ModelBase.register("Gemma3nForCausalLM", "Gemma3nForConditionalGeneration") @ModelBase.register("Gemma3nForCausalLM", "Gemma3nForConditionalGeneration")
@@ -6253,7 +6253,7 @@ class Gemma3NModel(Gemma3Model):
# Continue with normal processing # Continue with normal processing
name = name.replace("language_model.", "") name = name.replace("language_model.", "")
yield from super().modify_tensors(data_torch, name, bid) yield from ModelBase.modify_tensors(self, data_torch, name, bid)
return return
if "altup_unembed_projections" in name: if "altup_unembed_projections" in name:
@@ -6270,7 +6270,7 @@ class Gemma3NModel(Gemma3Model):
raise ValueError(f"Unknown name: {name}") raise ValueError(f"Unknown name: {name}")
out = self._stack_matrices(self._altup_unembd) out = self._stack_matrices(self._altup_unembd)
if out is not None: if out is not None:
yield from super().modify_tensors(out, "model.altup_unembed_projections.weight", bid) yield from ModelBase.modify_tensors(self, out, "model.altup_unembed_projections.weight", bid)
return return
else: else:
return return
@@ -6287,7 +6287,7 @@ class Gemma3NModel(Gemma3Model):
raise ValueError(f"Unknown name: {name}") raise ValueError(f"Unknown name: {name}")
out = self._stack_matrices(self._altup_proj) out = self._stack_matrices(self._altup_proj)
if out is not None: if out is not None:
yield from super().modify_tensors(out, "model.altup_projections.weight", bid) yield from ModelBase.modify_tensors(self, out, "model.altup_projections.weight", bid)
return return
else: else:
return return
@@ -8803,8 +8803,8 @@ class GraniteMoeModel(GraniteModel):
ffn_dim = self.hparams["intermediate_size"] ffn_dim = self.hparams["intermediate_size"]
assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size" assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size"
gate, up = data_torch.split(ffn_dim, dim=-2) gate, up = data_torch.split(ffn_dim, dim=-2)
yield from super().modify_tensors(gate, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), bid) yield from ModelBase.modify_tensors(self, gate, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), bid)
yield from super().modify_tensors(up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), bid) yield from ModelBase.modify_tensors(self, up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), bid)
has_experts = bool(self.hparams.get('num_local_experts')) has_experts = bool(self.hparams.get('num_local_experts'))
@@ -8813,15 +8813,15 @@ class GraniteMoeModel(GraniteModel):
assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size" assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
gate, up = data_torch.split(ffn_dim, dim=-2) gate, up = data_torch.split(ffn_dim, dim=-2)
if has_experts: if has_experts:
yield from super().modify_tensors(gate,self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), bid) yield from ModelBase.modify_tensors(self, gate,self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), bid)
yield from super().modify_tensors(up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), bid) yield from ModelBase.modify_tensors(self, up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), bid)
return return
yield from super().modify_tensors(gate, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), bid) yield from ModelBase.modify_tensors(self, gate, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), bid)
yield from super().modify_tensors(up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), bid) yield from ModelBase.modify_tensors(self, up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), bid)
return return
if not has_experts and name.endswith("shared_mlp.output_linear.weight"): if not has_experts and name.endswith("shared_mlp.output_linear.weight"):
yield from super().modify_tensors(data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN, bid), bid) yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN, bid), bid)
return return
yield from super().modify_tensors(data_torch, name, bid) yield from super().modify_tensors(data_torch, name, bid)