Skip to content

[kimi25 rl part4] Support K25 HF weight conversion between BF16\FP8\INT4 #1757

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Gao016 wants to merge 1 commit into
base: main
Choose a base branch
from
+20 −6
Open

[kimi25 rl part4] Support K25 HF weight conversion between BF16\FP8\INT4 #1757

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions tools/convert_hf_to_fp8.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -134,6 +134,8 @@ 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
Expand Down
2 changes: 2 additions & 0 deletions tools/convert_hf_to_int4.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -77,6 +77,8 @@ def main():
"re:.*shared_experts.*",
"re:.*mlp\\.(gate|up|gate_up|down)_proj.*",
"re:.*mlp\\.gate\\.*",
"re:vision_tower.*",
"re:mm_projector.*",
]

recipe = GPTQModifier(
Expand Down
2 changes: 2 additions & 0 deletions tools/convert_hf_to_int4_direct.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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",
)
Expand Down
20 changes: 14 additions & 6 deletions tools/convert_k2_thinking_int4_to_bf16.py → tools/convert_kimi_int4_to_bf16.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -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:
Expand All @@ -20,7 +20,8 @@

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
Expand All @@ -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", {})
Expand Down Expand Up @@ -218,7 +220,13 @@ def main():
dst_path = os.path.join(output_dir, fname)
if fname == "model.safetensors.index.json":
continue
if fname.endswith(".json") or fname.endswith(".py") or fname.startswith("tokenizer"):
if (
fname.endswith(".json")
or fname.endswith(".py")
or fname.startswith("tokenizer")
or fname.endswith(".jinja")
or fname.endswith(".model")
):
shutil.copy2(src_path, dst_path)

# Generate new index
Expand Down