happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

convert: simplify merge tensor condition

This commit is contained in:
Aman Gupta 2026-01-28 10:35:28 +01:00
parent d527830137
commit 566884c5dd
1 changed files with 15 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

35
convert_hf_to_gguf.py
View File

@ -524,28 +524,23 @@ class ModelBase:
if self.fuse_gate_up_exps and bid is not None:
if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_GATE_EXP, bid):
self._gate_exp_buffer[bid] = data_torch
# Check if up_exps is already buffered for this layer
if bid in self._up_exp_buffer:
gate_data = self._gate_exp_buffer.pop(bid)
up_data = self._up_exp_buffer.pop(bid)
# gate/up shape: (n_expert, n_ff, n_embd), concatenate to (n_expert, n_ff*2, n_embd)
fused_data = torch.cat([gate_data, up_data], dim=1)
fused_name = self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_UP_EXP, bid)
logger.info(f"Fused gate_exps and up_exps for layer {bid}")
return [(fused_name, fused_data)]
return [] # Wait for up_exps
elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_UP_EXP, bid):
self._up_exp_buffer[bid] = data_torch
# Check if gate_exps is already buffered for this layer
if bid in self._gate_exp_buffer:
gate_data = self._gate_exp_buffer.pop(bid)
up_data = self._up_exp_buffer.pop(bid)
# gate/up shape: (n_expert, n_ff, n_embd), concatenate to (n_expert, n_ff*2, n_embd)
fused_data = torch.cat([gate_data, up_data], dim=1)
fused_name = self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_UP_EXP, bid)
logger.info(f"Fused gate_exps and up_exps for layer {bid}")
return [(fused_name, fused_data)]
return [] # Wait for gate_exps
# Check if both gate and up are buffered for this layer
if bid in self._gate_exp_buffer and bid in self._up_exp_buffer:
gate_data = self._gate_exp_buffer.pop(bid)
up_data = self._up_exp_buffer.pop(bid)
# gate/up shape: (n_expert, n_ff, n_embd), concatenate to (n_expert, n_ff*2, n_embd)
fused_data = torch.cat([gate_data, up_data], dim=1)
fused_name = self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_UP_EXP, bid)
logger.info(f"Fused gate_exps and up_exps for layer {bid}")
return [(fused_name, fused_data)]
# If we buffered a gate/up tensor, wait for the other
if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_GATE_EXP, bid) or \
self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_UP_EXP, bid):
return []
return [(new_name, data_torch)]