File size: 3,248 Bytes
c7ab596 db50a98 c7ab596 db50a98 c7ab596 db50a98 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | # Patch 1: speculative.py
spec_path = "/usr/local/lib/python3.12/dist-packages/vllm/config/speculative.py"
with open(spec_path) as f:
c = f.read()
old = '"deepseek_v3", "deepseek_v32", "glm_moe_dsa"'
new = '"deepseek_v3", "deepseek_v32", "glm_moe_dsa", "kimi_k25"'
if old in c and "kimi_k25" not in c:
c = c.replace(old, new)
with open(spec_path, "w") as f: f.write(c)
print("speculative.py PATCHED")
else:
print("speculative.py already patched")
# Patch 2: kimi_k25.py
k25_path = "/usr/local/lib/python3.12/dist-packages/vllm/model_executor/models/kimi_k25.py"
with open(k25_path) as f:
c = f.read()
if '"model.layers.": "language_model.model.layers."' not in c:
c = c.replace(
'"language_model.layers.": "language_model.model.layers.",',
'"language_model.layers.": "language_model.model.layers.",\n "model.layers.": "language_model.model.layers.",')
with open(k25_path, "w") as f: f.write(c)
print("kimi_k25.py PATCHED")
else:
print("kimi_k25.py already patched")
# Patch 3: deepseek_mtp.py - extract text_config
mtp_path = "/usr/local/lib/python3.12/dist-packages/vllm/model_executor/models/deepseek_mtp.py"
with open(mtp_path) as f:
c = f.read()
if "text_config" not in c:
import_marker = "from .utils import maybe_prefix"
helper = '''from .utils import maybe_prefix
def _get_text_config(hf_config):
"""Extract text_config from VLM configs (e.g. KimiK25Config)."""
return getattr(hf_config, 'text_config', hf_config)'''
c = c.replace(import_marker, helper)
c = c.replace(
"config = vllm_config.model_config.hf_config\n self.mtp_start_layer_idx",
"config = _get_text_config(vllm_config.model_config.hf_config)\n self.mtp_start_layer_idx")
c = c.replace(
"self.config = vllm_config.model_config.hf_config\n self.model = DeepSeekMultiTokenPredictor",
"self.config = _get_text_config(vllm_config.model_config.hf_config)\n self.model = DeepSeekMultiTokenPredictor")
c = c.replace(
"config = vllm_config.speculative_config.draft_model_config.hf_config\n self.config = config",
"config = _get_text_config(vllm_config.speculative_config.draft_model_config.hf_config)\n self.config = config")
with open(mtp_path, "w") as f: f.write(c)
print("deepseek_mtp.py PATCHED")
else:
print("deepseek_mtp.py already patched")
# Patch 4: eagle.py - handle KimiK25Config missing image_token_index
eagle_path = "/usr/local/lib/python3.12/dist-packages/vllm/v1/spec_decode/eagle.py"
with open(eagle_path) as f:
c = f.read()
old_eagle = " self.model.config.image_token_index = (\n target_model.config.image_token_index\n )"
new_eagle = " self.model.config.image_token_index = getattr(\n target_model.config, 'image_token_index',\n getattr(target_model.config, 'media_placeholder_token_id', 0)\n )"
if "media_placeholder_token_id" not in c:
c = c.replace(old_eagle, new_eagle)
with open(eagle_path, "w") as f: f.write(c)
print("eagle.py PATCHED (KimiK25 image_token_index fallback)")
else:
print("eagle.py already patched")
|