[Bugfix][Hardware][AMD] Consolidate FP8 min/max values into helper function

Add get_fp8_min_max() helper in quant_utils.py to centralize the
FP8 min/max value logic for ROCm fnuz dtype handling.

On ROCm with torch.float8_e4m3fnuz, using PyTorch's default finfo.max
(240.0) causes accuracy issues with dynamic quantization. The correct
value is 224.0 for fnuz dtype.

This change:
- Adds get_fp8_min_max(dtype) helper returning (fp8_min, fp8_max) tuple
- Updates input_quant_fp8.py to use the helper
- Updates fp8_utils.py per_token_group_quant_fp8() to use the helper
- Updates deep_gemm.py per_block_cast_to_fp8() to use the helper
- Updates tests/kernels/quant_utils.py to use the helper

Fixes #30360

Signed-off-by: c0de128 <kevin.mckay@outlook.com>

tests/kernels/quant_utils.py

@@ -4,13 +4,13 @@
 
 import torch
 
-from vllm.model_executor.layers.quantization.utils.quant_utils import group_broadcast
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    get_fp8_min_max,
+    group_broadcast,
+)
 from vllm.platforms import current_platform
 from vllm.utils.math_utils import round_up
 
-# Using the default value (240.0) from pytorch will cause accuracy
-# issue on dynamic quantization models. Here use 224.0 for rocm.
-ROCM_FP8FNUZ_MAX = 224.0
 FP8_DTYPE = current_platform.fp8_dtype()
 
 
@@ -25,16 +25,12 @@ def ref_dynamic_per_token_quant(
     if scale_ub is not None:
         assert quant_dtype == FP8_DTYPE
 
-    qtype_traits = (
-        torch.iinfo(quant_dtype)
-        if quant_dtype == torch.int8
-        else torch.finfo(quant_dtype)
-    )
-    use_fp8fnuz = (
-        current_platform.is_fp8_fnuz() and quant_dtype == current_platform.fp8_dtype()
-    )
-    qtype_traits_max = ROCM_FP8FNUZ_MAX if use_fp8fnuz else qtype_traits.max
-    qtype_traits_min = -ROCM_FP8FNUZ_MAX if use_fp8fnuz else qtype_traits.min
+    if quant_dtype == torch.int8:
+        qtype_traits = torch.iinfo(quant_dtype)
+        qtype_traits_min = qtype_traits.min
+        qtype_traits_max = qtype_traits.max
+    else:
+        qtype_traits_min, qtype_traits_max = get_fp8_min_max(quant_dtype)
     qtype_max = as_float32_tensor(qtype_traits_max)
     s_1 = as_float32_tensor(1.0)
     s_512 = as_float32_tensor(512.0)
@@ -72,17 +68,7 @@ def ref_dynamic_per_token_quant(
 def ref_dynamic_per_tensor_fp8_quant(
     x: torch.Tensor,
 ) -> tuple[torch.Tensor, torch.Tensor]:
-    fp8_traits = torch.finfo(FP8_DTYPE)
-    fp8_traits_max = (
-        ROCM_FP8FNUZ_MAX
-        if current_platform.is_rocm() and current_platform.is_fp8_fnuz()
-        else fp8_traits.max
-    )
-    fp8_traits_min = (
-        -ROCM_FP8FNUZ_MAX
-        if current_platform.is_rocm() and current_platform.is_fp8_fnuz()
-        else fp8_traits.min
-    )
+    fp8_traits_min, fp8_traits_max = get_fp8_min_max(FP8_DTYPE)
     fp8_max = as_float32_tensor(fp8_traits_max)
     one = as_float32_tensor(1.0)
 

vllm/model_executor/layers/quantization/input_quant_fp8.py

@@ -7,15 +7,14 @@ import torch.nn.functional as F
 from vllm import _custom_ops as ops
 from vllm._aiter_ops import rocm_aiter_ops
 from vllm.model_executor.custom_op import CustomOp
-from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    GroupShape,
+    get_fp8_min_max,
+)
 from vllm.platforms import current_platform
 
-# Using the default value (240.0) from pytorch will cause accuracy
-# issue on dynamic quantization models. Here use 224.0 for fnuz on ROCm.
 _FP8_DTYPE = current_platform.fp8_dtype()
-_FP8_FINFO = torch.finfo(_FP8_DTYPE)
-_FP8_MAX = 224.0 if current_platform.is_fp8_fnuz() else _FP8_FINFO.max
-_FP8_MIN = -224.0 if current_platform.is_fp8_fnuz() else _FP8_FINFO.min
+_FP8_MIN, _FP8_MAX = get_fp8_min_max(_FP8_DTYPE)
 _FP8_MIN_SCALING_FACTOR = 1.0 / (_FP8_MAX * 512.0)
 
 

vllm/model_executor/layers/quantization/utils/fp8_utils.py

@@ -15,7 +15,10 @@ from vllm import _custom_ops as ops
 from vllm._aiter_ops import rocm_aiter_ops
 from vllm.logger import init_logger
 from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
-from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    GroupShape,
+    get_fp8_min_max,
+)
 from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
     CUTLASS_BLOCK_FP8_SUPPORTED,
 )
@@ -748,12 +751,7 @@ def per_token_group_quant_fp8(
     )
     assert x.stride(-1) == 1, "`x` groups must be contiguous"
 
-    # Using the default value (240.0) from pytorch will cause accuracy
-    # issue on dynamic quantization models. Here use 224.0 for fnuz on ROCm
-    # platforms that use the torch.float8_e4mefnuz dtype.
-    finfo = torch.finfo(dtype)
-    fp8_min = -224.0 if current_platform.is_fp8_fnuz() else finfo.min
-    fp8_max = 224.0 if current_platform.is_fp8_fnuz() else finfo.max
+    fp8_min, fp8_max = get_fp8_min_max(dtype)
 
     assert out_q is None or out_q.shape == x.shape
     x_q = out_q

vllm/model_executor/layers/quantization/utils/quant_utils.py

@@ -19,6 +19,28 @@ FP8_DTYPE = current_platform.fp8_dtype()
 FP4_DTYPE = torch.uint8
 
 
+def get_fp8_min_max(dtype: torch.dtype | None = None) -> tuple[float, float]:
+    """
+    Get the min and max values for FP8 quantization.
+
+    On ROCm with torch.float8_e4m3fnuz (fnuz), the default PyTorch finfo.max
+    (240.0) causes accuracy issues with dynamic quantization models.
+    Use 224.0 instead for fnuz dtype.
+
+    Args:
+        dtype: FP8 dtype (defaults to platform's FP8 dtype if None)
+
+    Returns:
+        Tuple of (fp8_min, fp8_max) values
+    """
+    if dtype is None:
+        dtype = FP8_DTYPE
+    finfo = torch.finfo(dtype)
+    if current_platform.is_fp8_fnuz():
+        return -224.0, 224.0
+    return finfo.min, finfo.max
+
+
 # Use proxy as NamedTuple direct subclasses cannot have static members
 class _GroupShape(NamedTuple):
     row: int

vllm/utils/deep_gemm.py

@@ -16,6 +16,9 @@ import torch
 
 import vllm.envs as envs
 from vllm.logger import logger
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    get_fp8_min_max,
+)
 from vllm.platforms import current_platform
 from vllm.utils.import_utils import has_deep_gemm
 from vllm.utils.math_utils import cdiv
@@ -355,7 +358,8 @@ def per_block_cast_to_fp8(
     x_padded[:m, :n] = x
     x_view = x_padded.view(-1, block_m, x_padded.size(1) // block_n, block_n)
     x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
-    sf = x_amax / 224.0 if current_platform.is_fp8_fnuz() else x_amax / 448.0
+    _, fp8_max = get_fp8_min_max(fp8_dtype)
+    sf = x_amax / fp8_max
     sf = _ceil_to_ue8m0(sf) if use_ue8m0 else sf
     x_scaled = (x_view * (1.0 / sf)).to(fp8_dtype)
     return x_scaled.view_as(x_padded)[:m, :n].contiguous(), sf.view(