tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-28 04:15:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix 32B model

Browse Source
This commit is contained in:
Xuan Son Nguyen
2025-04-30 22:04:06 +02:00
parent 474933e252
commit 651752f1ae
1 changed files with 14 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
convert_hf_to_gguf.py
View File

@ -2579,15 +2579,20 @@ class Qwen2VLVisionModel(VisionModel):
elif self.global_config['model_type'] == 'qwen2_5_vl': elif self.global_config['model_type'] == 'qwen2_5_vl':
self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN25VL) self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN25VL)
self.gguf_writer.add_vision_use_silu(True) self.gguf_writer.add_vision_use_silu(True)
# find n_wa_pattern (window attention pattern) out_hidden_size = hparams.get("out_hidden_size")
fullatt_block_indexes = hparams.get("fullatt_block_indexes") if out_hidden_size == 5120:
assert fullatt_block_indexes is not None, "fullatt_block_indexes is required for qwen2_5_vl" # 32B model does not have n_wa_pattern, the other models do
n_wa_pattern = fullatt_block_indexes[0] + 1 self.gguf_writer.add_vision_n_wa_pattern(0)
# validate n_wa_pattern else:
for i in range(1, len(fullatt_block_indexes)): # find n_wa_pattern (window attention pattern)
if fullatt_block_indexes[i] - fullatt_block_indexes[i - 1] != n_wa_pattern: fullatt_block_indexes = hparams.get("fullatt_block_indexes")
raise ValueError(f"Invalid fullatt_block_indexes: {fullatt_block_indexes}") assert fullatt_block_indexes is not None, "fullatt_block_indexes is required for qwen2_5_vl"
self.gguf_writer.add_vision_n_wa_pattern(n_wa_pattern) n_wa_pattern = fullatt_block_indexes[0] + 1
# validate n_wa_pattern
for i in range(1, len(fullatt_block_indexes)):
if fullatt_block_indexes[i] - fullatt_block_indexes[i - 1] != n_wa_pattern:
raise ValueError(f"Invalid fullatt_block_indexes: {fullatt_block_indexes}")
self.gguf_writer.add_vision_n_wa_pattern(n_wa_pattern)
else: else:
raise ValueError(f"Unknown QwenVL model type: {self.global_config['model_type']}") raise ValueError(f"Unknown QwenVL model type: {self.global_config['model_type']}")
# default values below are taken from HF tranformers code # default values below are taken from HF tranformers code