[quant] fix fp32 downcasting (#19844)

Co-authored-by: qiuxuan.lzw <qiuxuan.lzw@alibaba-inc.com>

python/sglang/srt/environ.py

@@ -326,6 +326,7 @@ class Envs:
     SGLANG_NVFP4_CKPT_FP8_GEMM_IN_ATTN = EnvBool(False)
     SGLANG_PER_TOKEN_GROUP_QUANT_8BIT_V2 = EnvBool(False)
     SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE = EnvBool(False)
+    SGLANG_QUANT_ALLOW_DOWNCASTING = EnvBool(False)
 
     # Flashinfer
     SGLANG_IS_FLASHINFER_AVAILABLE = EnvBool(True)

python/sglang/srt/layers/parameter.py

@@ -7,6 +7,7 @@ from typing import Callable, Optional, Union
 import torch
 from torch.nn import Parameter
 
+from sglang.srt.environ import envs
 from sglang.srt.layers.utils import pad_or_narrow_weight
 from sglang.srt.utils import is_cpu
 
@@ -65,7 +66,7 @@ def copy_with_check(target: torch.Tensor, loaded_weight: torch.Tensor):
         raise ValueError(
             f"Unsupported copy between dtypes: {target.dtype=}, {loaded_weight.dtype=}"
         )
-    if target_rank < loaded_rank:
+    if target_rank < loaded_rank and not envs.SGLANG_QUANT_ALLOW_DOWNCASTING.get():
         raise ValueError(
             f"Downcasting not allowed: {target.dtype=}, {loaded_weight.dtype=}"
         )