Fix fp8_cast_bf16.py: https://github.com/deepseek-ai/DeepSeek-V3/commit/8f1c9488b53068992f9525fab03b1868e6f7c8c1
Browse files- inference/fp8_cast_bf16.py +37 -11
inference/fp8_cast_bf16.py
CHANGED
@@ -16,32 +16,58 @@ def main(fp8_path, bf16_path):
|
|
16 |
with open(model_index_file, "r") as f:
|
17 |
model_index = json.load(f)
|
18 |
weight_map = model_index["weight_map"]
|
19 |
-
fp8_weight_names = []
|
20 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
safetensor_files = list(glob(os.path.join(fp8_path, "*.safetensors")))
|
|
|
22 |
for safetensor_file in tqdm(safetensor_files):
|
23 |
file_name = os.path.basename(safetensor_file)
|
24 |
-
|
|
|
|
|
25 |
new_state_dict = {}
|
26 |
-
for weight_name, weight in
|
27 |
if weight_name.endswith("_scale_inv"):
|
28 |
continue
|
29 |
-
elif weight.element_size() == 1:
|
30 |
scale_inv_name = f"{weight_name}_scale_inv"
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
|
|
|
|
|
|
|
|
35 |
else:
|
36 |
new_state_dict[weight_name] = weight
|
|
|
37 |
new_safetensor_file = os.path.join(bf16_path, file_name)
|
38 |
save_file(new_state_dict, new_safetensor_file)
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
|
|
|
40 |
new_model_index_file = os.path.join(bf16_path, "model.safetensors.index.json")
|
41 |
for weight_name in fp8_weight_names:
|
42 |
scale_inv_name = f"{weight_name}_scale_inv"
|
43 |
-
|
44 |
-
|
45 |
with open(new_model_index_file, "w") as f:
|
46 |
json.dump({"metadata": {}, "weight_map": weight_map}, f, indent=2)
|
47 |
|
@@ -52,4 +78,4 @@ if __name__ == "__main__":
|
|
52 |
parser.add_argument("--output-bf16-hf-path", type=str, required=True)
|
53 |
args = parser.parse_args()
|
54 |
main(args.input_fp8_hf_path, args.output_bf16_hf_path)
|
55 |
-
|
|
|
16 |
with open(model_index_file, "r") as f:
|
17 |
model_index = json.load(f)
|
18 |
weight_map = model_index["weight_map"]
|
|
|
19 |
|
20 |
+
# Cache for loaded safetensor files
|
21 |
+
loaded_files = {}
|
22 |
+
fp8_weight_names = []
|
23 |
+
|
24 |
+
# Helper function to get tensor from the correct file
|
25 |
+
def get_tensor(tensor_name):
|
26 |
+
file_name = weight_map[tensor_name]
|
27 |
+
if file_name not in loaded_files:
|
28 |
+
file_path = os.path.join(fp8_path, file_name)
|
29 |
+
loaded_files[file_name] = load_file(file_path, device="cuda")
|
30 |
+
return loaded_files[file_name][tensor_name]
|
31 |
+
|
32 |
safetensor_files = list(glob(os.path.join(fp8_path, "*.safetensors")))
|
33 |
+
safetensor_files.sort()
|
34 |
for safetensor_file in tqdm(safetensor_files):
|
35 |
file_name = os.path.basename(safetensor_file)
|
36 |
+
current_state_dict = load_file(safetensor_file, device="cuda")
|
37 |
+
loaded_files[file_name] = current_state_dict
|
38 |
+
|
39 |
new_state_dict = {}
|
40 |
+
for weight_name, weight in current_state_dict.items():
|
41 |
if weight_name.endswith("_scale_inv"):
|
42 |
continue
|
43 |
+
elif weight.element_size() == 1: # FP8 weight
|
44 |
scale_inv_name = f"{weight_name}_scale_inv"
|
45 |
+
try:
|
46 |
+
# Get scale_inv from the correct file
|
47 |
+
scale_inv = get_tensor(scale_inv_name)
|
48 |
+
fp8_weight_names.append(weight_name)
|
49 |
+
new_state_dict[weight_name] = weight_dequant(weight, scale_inv)
|
50 |
+
except KeyError:
|
51 |
+
print(f"Warning: Missing scale_inv tensor for {weight_name}, skipping conversion")
|
52 |
+
new_state_dict[weight_name] = weight
|
53 |
else:
|
54 |
new_state_dict[weight_name] = weight
|
55 |
+
|
56 |
new_safetensor_file = os.path.join(bf16_path, file_name)
|
57 |
save_file(new_state_dict, new_safetensor_file)
|
58 |
+
|
59 |
+
# Memory management: keep only the 2 most recently used files
|
60 |
+
if len(loaded_files) > 2:
|
61 |
+
oldest_file = next(iter(loaded_files))
|
62 |
+
del loaded_files[oldest_file]
|
63 |
+
torch.cuda.empty_cache()
|
64 |
|
65 |
+
# Update model index
|
66 |
new_model_index_file = os.path.join(bf16_path, "model.safetensors.index.json")
|
67 |
for weight_name in fp8_weight_names:
|
68 |
scale_inv_name = f"{weight_name}_scale_inv"
|
69 |
+
if scale_inv_name in weight_map:
|
70 |
+
weight_map.pop(scale_inv_name)
|
71 |
with open(new_model_index_file, "w") as f:
|
72 |
json.dump({"metadata": {}, "weight_map": weight_map}, f, indent=2)
|
73 |
|
|
|
78 |
parser.add_argument("--output-bf16-hf-path", type=str, required=True)
|
79 |
args = parser.parse_args()
|
80 |
main(args.input_fp8_hf_path, args.output_bf16_hf_path)
|
81 |
+
|