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

remove spammy shit

Browse Source
This commit is contained in:
biondizzle
2026-05-12 20:57:04 +00:00
parent f0652693a6
commit d68e113af1
1 changed files with 1 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
patches/deepseek_v4.py
View File

@@ -669,14 +669,6 @@ class DeepseekV4MegaMoEExperts(nn.Module):
self._check_runtime_supported()
from deep_gemm.mega import transform_nvfp4_weights_for_mega_moe
# Debug: log shapes/dtypes before transform
print(f"[NVFP4-MoE] w13_weight: shape={self.w13_weight.shape} dtype={self.w13_weight.dtype}")
print(f"[NVFP4-MoE] w13_weight_scale: shape={self.w13_weight_scale.shape} dtype={self.w13_weight_scale.dtype}")
print(f"[NVFP4-MoE] w13_weight_scale_2: shape={self.w13_weight_scale_2.shape} dtype={self.w13_weight_scale_2.dtype}")
print(f"[NVFP4-MoE] w2_weight: shape={self.w2_weight.shape} dtype={self.w2_weight.dtype}")
print(f"[NVFP4-MoE] w2_weight_scale: shape={self.w2_weight_scale.shape} dtype={self.w2_weight_scale.dtype}")
print(f"[NVFP4-MoE] w2_weight_scale_2: shape={self.w2_weight_scale_2.shape} dtype={self.w2_weight_scale_2.dtype}")
# === Native NVFP4 path ===
# The DeepGEMM nvfp4 mega_moe kernel consumes NVFP4 directly:
# - E2M1 packed uint8 (same as checkpoint)
@@ -2082,11 +2074,7 @@ class DeepseekV4Model(nn.Module):
delattr(mod, attr)
from vllm.model_executor.layers.linear import UnquantizedLinearMethod
mod.quant_method = UnquantizedLinearMethod()
# Safety check: UnquantizedLinearMethod with FP8 weight will crash CUBLAS
if mod.weight.dtype in (torch.float8_e4m3fn, torch.float8_e5m2):
print(f"[NVFP4-FP8-CRASH] {getattr(mod, 'prefix', 'unknown')}: "
f"weight is {mod.weight.dtype} but quant_method=UnquantizedLinearMethod! "
f"This will crash CUBLAS.")
for attr in ("weight_scale", "weight_scale_2", "input_scale"):
if hasattr(mod, attr):
delattr(mod, attr)
@@ -2192,8 +2180,6 @@ class DeepseekV4Model(nn.Module):
# fused_wkv_wgate.weight = cat([wkv, wgate], dim=0) → (2*head_dim, hidden_size)
w_fused = torch.cat([wkv_dequant, wgate_dequant], dim=0)
# DEBUG: log shapes to diagnose compressor weight mismatch
print(f"NVFP4 compressor layer {layer_idx}: wkv={wkv_dequant.shape}, wgate={wgate_dequant.shape}, fused={w_fused.shape}, existing_param={fused_mod.weight.shape}")
# Replace the weight
fused_mod.weight = torch.nn.Parameter(w_fused, requires_grad=False)