biondizzle/deepseek-v4-quant
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Remove ModuleList patch (already fixed in modelopt 0.45), fix numbering

Browse Source
This commit is contained in:
biondizzle
2026-05-09 06:10:18 +00:00
parent 382c1d872f
commit 3907838409
1 changed files with 3 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
scripts/quantize_nvfp4.py
View File

@@ -48,23 +48,9 @@ CALIB_SAVE_PATH = "/root/nvidia-meeting/v4_nvfp4_calibrated_state.pt"
def apply_patches():
"""Apply runtime patches for V4 compatibility."""
from modelopt.torch.quantization.nn.modules import quant_module, tensor_quantizer as tq_module
from modelopt.torch.quantization.nn.modules import tensor_quantizer as tq_module
# 1. Patch quant_module.py for V4's ModuleList expert quantizers
orig_iter = quant_module._QuantFusedExperts.iter_weights_for_calibration
def patched_iter_weights_for_calibration(self, **kwargs):
"""Handle V4's nn.ModuleList expert quantizers (vs singular TensorQuantizer)."""
for name, quantizer in self.named_modules():
if not isinstance(quantizer, quant_module.TensorQuantizer):
continue
if quantizer.is_enabled:
yield name, quantizer
quant_module._QuantFusedExperts.iter_weights_for_calibration = patched_iter_weights_for_calibration
print("✓ Patched _QuantFusedExperts.iter_weights_for_calibration for V4 ModuleList")
# 2. Patch TensorQuantizer.export_amax to move _amax to CPU before reading
# 1. Patch TensorQuantizer.export_amax to move _amax to CPU before reading
orig_export_amax = tq_module.TensorQuantizer.export_amax
def patched_export_amax(self):
@@ -77,7 +63,7 @@ def apply_patches():
tq_module.TensorQuantizer.export_amax = patched_export_amax
print("✓ Patched TensorQuantizer.export_amax (CPU safety)")
# 3. Patch NVFP4QTensor.get_activation_scaling_factor for graceful degradation
# 2. Patch NVFP4QTensor.get_activation_scaling_factor for graceful degradation
from modelopt.torch.quantization.qtensor import nvfp4_tensor
orig_get_asf = nvfp4_tensor.NVFP4QTensor.get_activation_scaling_factor