biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: remove prefix_filter from layer tensor loading

Browse Source
This commit is contained in:
biondizzle
2026-05-16 02:15:55 +00:00
parent 303b6a8993
commit ea5ee7c1f7
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/layertest.py
View File

@@ -396,7 +396,7 @@ def main():
print(" STEP 1: Loading original MXFP4 checkpoint")
print("="*70)
orig_tensors = load_layer_tensors(ORIG_MODEL_DIR, LAYER_IDX, prefix_filter="experts")
orig_tensors = load_layer_tensors(ORIG_MODEL_DIR, LAYER_IDX)
print_layer_keys(orig_tensors, "Original checkpoint (MXFP4)")
# Dequantize to BF16
@@ -430,7 +430,7 @@ def main():
print(" STEP 3: Loading NVFP4 checkpoint")
print("="*70)
nvfp4_tensors = load_layer_tensors(NVFP4_MODEL_DIR, LAYER_IDX, prefix_filter="experts")
nvfp4_tensors = load_layer_tensors(NVFP4_MODEL_DIR, LAYER_IDX)
print_layer_keys(nvfp4_tensors, "NVFP4 checkpoint")
# Verify dtype of weight_scale (should be float8_e4m3fn, NOT float8_e8m0fnu)