[mxfp] handle values close to max correctly w/o overflow

python/triton_kernels/tests/test_mxfp.py

@@ -45,6 +45,22 @@ def test_mxfp4_rounding_cases(dst_dtype, device):
     assert_equal(dequant_torch, dequant)
 
 
+@pytest.mark.parametrize("src_dtype", ["float4_e2m1", "float8_e5m2", "float8_e4m3fn"])
+@pytest.mark.parametrize("dst_dtype", ["float16", "bfloat16", "float32"])
+def test_mxfp_extreme_values(src_dtype, dst_dtype, device):
+    if "float8" in src_dtype and (is_cuda() and torch.cuda.get_device_capability()[0] < 9):
+        pytest.skip("Float8 not tested on A100")
+    src_dtype = dtype_str_to_torch(src_dtype)
+    dst_dtype = dtype_str_to_torch(dst_dtype)
+    BIG_VALUE = 65470 if dst_dtype == torch.float16 else 3.3895e38
+    x = torch.tensor([BIG_VALUE, BIG_VALUE], dtype=dst_dtype, device=device)
+    xq_value, xq_scale = downcast_to_mxfp(x, src_dtype, axis=-1)
+    xdq = upcast_from_mxfp(xq_value, xq_scale, dst_dtype, axis=-1)
+    xdq_ref = upcast_from_mxfp_torch(xq_value, xq_scale, dst_dtype, axis=-1)
+    assert_equal(xdq_ref, xdq)
+    assert not xdq.isinf().any()
+
+
 @pytest.mark.parametrize("src_dtype", ["float4_e2m1", "float8_e5m2", "float8_e4m3fn"])
 @pytest.mark.parametrize("dst_dtype", ["float16", "bfloat16", "float32"])
 def test_mxfp_quant_dequant(src_dtype, dst_dtype, device):

python/triton_kernels/triton_kernels/numerics_details/mxfp.py

@@ -297,6 +297,17 @@ def upcast_from_mxfp_torch(tensor: torch.Tensor, scale: torch.Tensor, target_dty
     padded_tensor = padded_tensor.view(*new_shape)
     dq_scale_padded = dq_scale.unsqueeze(-1)  # shape: [..., ceil(axis_shape/32), 1]
     out_padded = padded_tensor * dq_scale_padded
+    # Need to clamp since due to rounding, we can have overflow that was within
+    # the range before quantization.
+    # e.g., 3.3895e+38 -> log2(3.3895e+38 / max_fp8e4m3=448) ~= 119.17 -> round
+    # up to 120 + exp_bias=127 -> scale=247
+    # 3.3895e+38 / 2**120 ~= 254.9976 -> round to 256 in fp8e4m3fn
+    # Dequantization: 256 * 2**120 > 3.4e38 overflowing 3.38953139e38
+    finfo = torch.finfo(target_dtype)
+    out_padded = (padded_tensor * dq_scale_padded).clamp(finfo.min, finfo.max)
+    if tensor.dtype == torch.float8_e5m2:
+        # fp8e5m2 can have inf and we want to preserve so separately handle
+        out_padded = out_padded.where(~padded_tensor.isinf(), padded_tensor.to(target_dtype))
 
     # Flatten back and remove the padded tail
     out_padded = out_padded.view(*fp32_tensor.shape[:-1], new_axis_shape)

python/triton_kernels/triton_kernels/numerics_details/mxfp_details/_upcast_from_mxfp.py

@@ -119,6 +119,16 @@ def _upcast_from_mxfp(out_ptr, stride_o_outer, stride_o_quant: tl.constexpr, mx_
     scale = scale.reshape(dst_scale.shape)
 
     out_tensor = dst_tensor * dst_scale
+    if dst_dtype == tl.float32:
+        max_fin = 3.4028234663852886e+38
+    elif dst_dtype == tl.bfloat16:
+        max_fin = 3.3895313892515355e+38
+    else:
+        tl.static_assert(dst_dtype == tl.float16)
+        max_fin = 65504
+    # TODO: handle infinity same as upcast_from_mxfp_torch together with the
+    # above FIXME
+    out_tensor = tl.clamp(out_tensor, min=-max_fin, max=max_fin)
     # Correct any NaNs encoded via the scale.
     out_tensor = tl.where(scale == 0xFF, float("nan"), out_tensor)
     out_tensor = out_tensor.reshape([BLOCK_SIZE_OUT_DIM, BLOCK_SIZE_QUANT_DIM])