2/5 support kimi 2.5 full + lora: VL-aware quantization/conversion tools (#1220)

Co-authored-by: JiLi <22428217+GeLee-Q@users.noreply.github.com>

tools/convert_hf_to_fp8.py

@@ -137,6 +137,8 @@ def process_file(input_path, output_path, filename, strategy, block_size, result
             and "lm_head" not in key
             and "eh_proj" not in key
             and "weights_proj" not in key
+            and "vision_tower" not in key
+            and "mm_projector" not in key
         ):
             qw, s = quant_fp8(weights[key], strategy, block_size)
             q_weights[key] = qw

tools/convert_hf_to_int4.py

@@ -79,6 +79,8 @@ def main():
         "re:.*shared_experts.*",
         "re:.*mlp\\.(gate|up|gate_up|down)_proj.*",
         "re:.*mlp\\.gate\\.*",
+        "re:vision_tower.*",
+        "re:mm_projector.*",
     ]
 
     recipe = GPTQModifier(

tools/convert_hf_to_int4_direct.py

@@ -295,6 +295,8 @@ def parse_args():
             "re:.*shared_experts.*",
             "re:.*mlp\\.(gate|up|gate_up|down)_proj.*",
             "re:.*mlp\\.gate\\.*",
+            "re:vision_tower.*",
+            "re:mm_projector.*",
         ],
         help="Ignore Rules",
     )

tools/convert_kimi_int4_to_bf16.py

@@ -1,7 +1,7 @@
 """
 Usage:
 ------
-python convert_k2_thinking_int4_to_bf16.py [-h] --model-dir MODEL_DIR [--output-dir OUTPUT_DIR]
+python convert_kimi_int4_to_bf16.py [-h] --model-dir MODEL_DIR [--output-dir OUTPUT_DIR]
                            [--files FILE [FILE ...]] [--config-path CONFIG_PATH]
                            [--overwrite]
 options:
@@ -20,7 +20,8 @@ options:
 
 Example:
 --------
-python convert_k2_thinking_int4_to_bf16.py --model-dir /Kimi-K2-Thinking --output-dir /Kimi-K2-Thinking-bf16
+python convert_kimi_int4_to_bf16.py --model-dir /Kimi-K2-Thinking --output-dir /Kimi-K2-Thinking-bf16
+python convert_kimi_int4_to_bf16.py --model-dir /Kimi-K2.5 --output-dir /Kimi-K2.5-bf16
 """
 
 import argparse
@@ -40,10 +41,11 @@ def _load_config(model_dir: str, config_path: str | None) -> tuple[int, int, int
     cfg_path = config_path or os.path.join(model_dir, "config.json")
     with open(cfg_path) as f:
         cfg = json.load(f)
-    hidden_size = int(cfg.get("hidden_size"))
-    inter_size = int(cfg.get("moe_intermediate_size"))
+    loader = cfg.get("text_config", cfg)
+    hidden_size = int(loader.get("hidden_size"))
+    inter_size = int(loader.get("moe_intermediate_size"))
     group_size = int(
-        cfg.get("quantization_config", {})
+        loader.get("quantization_config", {})
         .get("config_groups", {})
         .get("group_0", {})
         .get("weights", {})
@@ -216,9 +218,11 @@ def main():
     for fname in os.listdir(model_dir):
         src_path = os.path.join(model_dir, fname)
         dst_path = os.path.join(output_dir, fname)
+        extensions_to_copy = (".json", ".py", ".jinja", ".model")
+
         if fname == "model.safetensors.index.json":
             continue
-        if fname.endswith(".json") or fname.endswith(".py") or fname.startswith("tokenizer"):
+        if fname.startswith("tokenizer") or any(fname.endswith(ext) for ext in extensions_to_copy):
             shutil.copy2(src_path, dst_path)
 
     # Generate new index