[Feature][Hardware][Amd] Add fp8 Linear Layer for Rocm (#7210)

csrc/quantization/fp8/common.cu

@@ -9,6 +9,18 @@
 
 #include "../../reduction_utils.cuh"
 
+#ifndef USE_ROCM
+using FP8_TYPE = c10::Float8_e4m3fn;
+C10_HOST_DEVICE constexpr auto FP8_E4M3_MAX =
+    std::numeric_limits<FP8_TYPE>::max();
+#else
+  #include "amd/hip_float8.h"
+using FP8_TYPE = c10::Float8_e4m3fnuz;
+// Using the default max value from pytorch (240.0) will cause accuracy
+// issue when running dynamic quantization. Here use 224.0f for rocm.
+constexpr auto FP8_E4M3_MAX = 224.0f;
+#endif
+
 namespace vllm {
 
 __device__ __forceinline__ float atomicMaxFloat(float* addr, float value) {
@@ -21,11 +33,9 @@ __device__ __forceinline__ float atomicMaxFloat(float* addr, float value) {
   return old;
 }
 
-#define FP8_E4M3_MAX std::numeric_limits<c10::Float8_e4m3fn>::max()
-
 template <bool is_scale_inverted>
-__device__ __forceinline__ c10::Float8_e4m3fn scaled_fp8_conversion(
+__device__ __forceinline__ FP8_TYPE scaled_fp8_conversion(float const val,
-    float const val, float const scale) {
+                                                          float const scale) {
   float x = 0.0f;
   if constexpr (is_scale_inverted) {
     x = val * scale;
@@ -34,7 +44,13 @@ __device__ __forceinline__ c10::Float8_e4m3fn scaled_fp8_conversion(
   }
 
   float r = fmax(-FP8_E4M3_MAX, fmin(x, FP8_E4M3_MAX));
+#ifndef USE_ROCM
   return static_cast<c10::Float8_e4m3fn>(r);
+#else
+  // Use hardware cvt instruction for fp8 on rocm
+  return c10::Float8_e4m3fnuz(hip_fp8(r).data,
+                              c10::Float8_e4m3fnuz::from_bits());
+#endif
 }
 
 // Compute the absolute maximum m of the input tensor and store
@@ -74,8 +90,7 @@ __global__ void segmented_max_reduction(float* __restrict__ scale,
   // Finally, since cache[0] contains the maximum for this thread block,
   // atomically write the max to the target location
   if (threadIdx.x == 0) {
-    atomicMaxFloat(scale,
+    atomicMaxFloat(scale, cache[0] / FP8_E4M3_MAX);
-                   cache[0] / std::numeric_limits<c10::Float8_e4m3fn>::max());
   }
 }
 
@@ -88,10 +103,10 @@ struct __align__(8) vec4_t {
 };
 
 typedef struct __align__(4) {
-  c10::Float8_e4m3fn x;
+  FP8_TYPE x;
-  c10::Float8_e4m3fn y;
+  FP8_TYPE y;
-  c10::Float8_e4m3fn z;
+  FP8_TYPE z;
-  c10::Float8_e4m3fn w;
+  FP8_TYPE w;
 }
 float8x4_t;
 
@@ -124,7 +139,7 @@ __device__ float thread_max_vec(scalar_t const* __restrict__ input,
 }
 
 template <typename scalar_t, bool is_scale_inverted>
-__device__ void scaled_fp8_conversion_vec(c10::Float8_e4m3fn* __restrict__ out,
+__device__ void scaled_fp8_conversion_vec(FP8_TYPE* __restrict__ out,
                                           scalar_t const* __restrict__ input,
                                           float const scale,
                                           int64_t const num_elems,
@@ -160,7 +175,7 @@ __device__ void scaled_fp8_conversion_vec(c10::Float8_e4m3fn* __restrict__ out,
 }
 
 template <typename scalar_t>
-__global__ void scaled_fp8_quant_kernel(c10::Float8_e4m3fn* __restrict__ out,
+__global__ void scaled_fp8_quant_kernel(FP8_TYPE* __restrict__ out,
                                         const scalar_t* __restrict__ input,
                                         const float* __restrict__ scale,
                                         int64_t num_elems) {
@@ -175,7 +190,7 @@ __global__ void scaled_fp8_quant_kernel(c10::Float8_e4m3fn* __restrict__ out,
 
 template <typename scalar_t>
 __global__ void dynamic_per_token_scaled_fp8_quant_kernel(
-    c10::Float8_e4m3fn* __restrict__ out, float* __restrict__ scale,
+    FP8_TYPE* __restrict__ out, float* __restrict__ scale,
     scalar_t const* __restrict__ input, float const* __restrict__ scale_ub,
     const int hidden_size) {
   float const min_scaling_factor = 1.0f / (FP8_E4M3_MAX * 512.f);
@@ -184,7 +199,7 @@ __global__ void dynamic_per_token_scaled_fp8_quant_kernel(
   int const token_idx = blockIdx.x;
 
   scalar_t const* __restrict__ token_input = &input[token_idx * hidden_size];
-  c10::Float8_e4m3fn* __restrict__ token_output = &out[token_idx * hidden_size];
+  FP8_TYPE* __restrict__ token_output = &out[token_idx * hidden_size];
 
   // For vectorization, token_input and token_output pointers need to be
   // aligned at 8-byte and 4-byte addresses respectively.
@@ -241,7 +256,7 @@ void static_scaled_fp8_quant(torch::Tensor& out,          // [..., d]
   VLLM_DISPATCH_FLOATING_TYPES(
       input.scalar_type(), "scaled_fp8_quant_kernel", [&] {
         vllm::scaled_fp8_quant_kernel<scalar_t><<<grid, block, 0, stream>>>(
-            out.data_ptr<c10::Float8_e4m3fn>(), input.data_ptr<scalar_t>(),
+            out.data_ptr<FP8_TYPE>(), input.data_ptr<scalar_t>(),
             scale.data_ptr<float>(), num_elems);
       });
 }
@@ -261,7 +276,7 @@ void dynamic_scaled_fp8_quant(torch::Tensor& out,          // [..., d]
         vllm::segmented_max_reduction<scalar_t><<<grid, block, 0, stream>>>(
             scale.data_ptr<float>(), input.data_ptr<scalar_t>(), num_elems);
         vllm::scaled_fp8_quant_kernel<scalar_t><<<grid, block, 0, stream>>>(
-            out.data_ptr<c10::Float8_e4m3fn>(), input.data_ptr<scalar_t>(),
+            out.data_ptr<FP8_TYPE>(), input.data_ptr<scalar_t>(),
             scale.data_ptr<float>(), num_elems);
       });
 }
@@ -284,7 +299,7 @@ void dynamic_per_token_scaled_fp8_quant(
       input.scalar_type(), "dynamic_per_token_scaled_fp8_quant_kernel", [&] {
         vllm::dynamic_per_token_scaled_fp8_quant_kernel<scalar_t>
             <<<grid, block, 0, stream>>>(
-                out.data_ptr<c10::Float8_e4m3fn>(), scales.data_ptr<float>(),
+                out.data_ptr<FP8_TYPE>(), scales.data_ptr<float>(),
                 input.data_ptr<scalar_t>(),
                 scale_ub.has_value() ? scale_ub->data_ptr<float>() : nullptr,
                 hidden_size);

tests/kernels/quant_utils.py

@@ -2,6 +2,13 @@ from typing import Optional, Tuple, Union
 
 import torch
 
+from vllm.utils import is_hip
+
+# Using the default value (240.0) from pytorch will cause accuracy
+# issue on dynamic quantization models. Here use 224.0 for rocm.
+ROCM_FP8_MAX = 224.0
+FP8_DTYPE = torch.float8_e4m3fnuz if is_hip() else torch.float8_e4m3fn
+
 
 def as_float32_tensor(x: Union[float, torch.tensor]) -> torch.tensor:
     return torch.as_tensor(x, dtype=torch.float32, device='cuda')
@@ -11,13 +18,15 @@ def ref_dynamic_per_token_quant(x: torch.tensor,
                                 scale_ub: Optional[torch.tensor] = None) \
         -> Tuple[torch.tensor, torch.tensor]:
 
-    assert quant_dtype in [torch.int8, torch.float8_e4m3fn]
+    assert quant_dtype in [torch.int8, FP8_DTYPE]
     if scale_ub is not None:
-        assert quant_dtype == torch.float8_e4m3fn
+        assert quant_dtype == FP8_DTYPE
 
     qtype_traits = torch.iinfo(quant_dtype) if quant_dtype == torch.int8 \
             else torch.finfo(quant_dtype)
-    qtype_max = as_float32_tensor(qtype_traits.max)
+    qtype_traits_max = ROCM_FP8_MAX if is_hip() else qtype_traits.max
+    qtype_traits_min = -ROCM_FP8_MAX if is_hip() else qtype_traits.min
+    qtype_max = as_float32_tensor(qtype_traits_max)
     s_1 = as_float32_tensor(1.0)
     s_512 = as_float32_tensor(512.0)
 
@@ -37,15 +46,15 @@ def ref_dynamic_per_token_quant(x: torch.tensor,
         iscales = as_float32_tensor(s_1 / scales)
         torch_out = as_float32_tensor(x) * iscales
         torch_out = torch_out.round()
-        torch_out = torch_out.clamp(qtype_traits.min,
+        torch_out = torch_out.clamp(qtype_traits_min,
-                                    qtype_traits.max).to(quant_dtype)
+                                    qtype_traits_max).to(quant_dtype)
     else:
-        assert quant_dtype == torch.float8_e4m3fn
+        assert quant_dtype == FP8_DTYPE
         min_scaling_factor = s_1 / (qtype_max * s_512)
         scales = scales.clamp(min=min_scaling_factor)
         torch_out = as_float32_tensor(x) / scales
-        torch_out = torch_out.clamp(qtype_traits.min,
+        torch_out = torch_out.clamp(qtype_traits_min,
-                                    qtype_traits.max).to(quant_dtype)
+                                    qtype_traits_max).to(quant_dtype)
 
     return torch_out, scales
 
@@ -56,8 +65,10 @@ def ref_dynamic_per_token_quant(x: torch.tensor,
 def ref_dynamic_per_tensor_fp8_quant(x: torch.tensor) \
                     -> Tuple[torch.tensor, torch.tensor]:
 
-    fp8_traits = torch.finfo(torch.float8_e4m3fn)
+    fp8_traits = torch.finfo(FP8_DTYPE)
-    fp8_max = as_float32_tensor(fp8_traits.max)
+    fp8_traits_max = ROCM_FP8_MAX if is_hip() else fp8_traits.max
+    fp8_traits_min = -ROCM_FP8_MAX if is_hip() else fp8_traits.min
+    fp8_max = as_float32_tensor(fp8_traits_max)
     one = as_float32_tensor(1.0)
 
     # For fp8, in order to match the cuda kernel output, we have to do exactly
@@ -68,5 +79,5 @@ def ref_dynamic_per_tensor_fp8_quant(x: torch.tensor) \
     ref_scale = x_max / fp8_max
     ref_iscale = one / ref_scale
     ref_out = (as_float32_tensor(x) * ref_iscale).clamp(
-        fp8_traits.min, fp8_traits.max).to(dtype=torch.float8_e4m3fn)
+        fp8_traits_min, fp8_traits_max).to(FP8_DTYPE)
     return ref_out, ref_scale.view((1, ))

tests/kernels/test_fp8_quant.py

@@ -2,7 +2,8 @@ import pytest
 import torch
 
 import vllm._custom_ops as ops
-from tests.kernels.quant_utils import (ref_dynamic_per_tensor_fp8_quant,
+from tests.kernels.quant_utils import (FP8_DTYPE,
+                                       ref_dynamic_per_tensor_fp8_quant,
                                        ref_dynamic_per_token_quant)
 
 DTYPES = [torch.half, torch.bfloat16, torch.float]
@@ -31,8 +32,7 @@ def test_dynamic_per_token_fp8_quant(num_tokens: int, hidden_size: int,
 
     scale_ub = torch.mean(x).to(dtype=torch.float32, device='cuda') \
             if scale_ub else None
-    ref_out, ref_scales = ref_dynamic_per_token_quant(x, torch.float8_e4m3fn,
+    ref_out, ref_scales = ref_dynamic_per_token_quant(x, FP8_DTYPE, scale_ub)
-                                                      scale_ub)
     ops_out, ops_scales = ops.scaled_fp8_quant(x,
                                                scale_ub=scale_ub,
                                                use_per_token_if_dynamic=True)

vllm/_custom_ops.py

@@ -369,9 +369,12 @@ def scaled_fp8_quant(
     # This code assumes batch_dim and num_tokens are flattened
     assert (input.ndim == 2)
     shape: Union[Tuple[int, int], torch.Size] = input.shape
+    # For rocm, the output fp8 dtype is torch.float_e3m3fnuz
+    out_dtype: torch.dtype = torch.float8_e4m3fnuz if vllm.utils.is_hip() \
+        else torch.float8_e4m3fn
     if num_token_padding:
         shape = (max(num_token_padding, input.shape[0]), shape[1])
-    output = torch.empty(shape, device=input.device, dtype=torch.float8_e4m3fn)
+    output = torch.empty(shape, device=input.device, dtype=out_dtype)
 
     if scale is None:
         if use_per_token_if_dynamic:

vllm/config.py

@@ -240,7 +240,7 @@ class ModelConfig:
 
     def _verify_quantization(self) -> None:
         supported_quantization = [*QUANTIZATION_METHODS]
-        rocm_supported_quantization = ["gptq", "squeezellm"]
+        rocm_supported_quantization = ["gptq", "squeezellm", "fp8"]
         optimized_quantization_methods = [
             "fp8", "marlin", "gptq_marlin_24", "gptq_marlin", "awq_marlin",
             "fbgemm_fp8", "compressed_tensors", "compressed-tensors"

vllm/model_executor/layers/quantization/fp8.py

@@ -20,10 +20,11 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
 from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
     all_close_1d, apply_fp8_linear, convert_to_channelwise,
     create_per_tensor_scale_param, cutlass_fp8_supported,
-    per_tensor_dequantize, requantize_with_max_scale)
+    normalize_e4m3fn_to_e4m3fnuz, per_tensor_dequantize,
+    requantize_with_max_scale)
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
-from vllm.utils import print_warning_once
+from vllm.utils import is_hip, print_warning_once
 
 ACTIVATION_SCHEMES = ["static", "dynamic"]
 
@@ -120,6 +121,9 @@ class Fp8LinearMethod(LinearMethodBase):
         capability = current_platform.get_device_capability()
         capability = capability[0] * 10 + capability[1]
         self.use_marlin = capability < 89 or envs.VLLM_TEST_FORCE_FP8_MARLIN
+        # Disable marlin for rocm
+        if is_hip():
+            self.use_marlin = False
 
     def create_weights(
         self,
@@ -168,6 +172,8 @@ class Fp8LinearMethod(LinearMethodBase):
                 scale = create_per_tensor_scale_param(output_partition_sizes,
                                                       **extra_weight_attrs)
                 layer.register_parameter("input_scale", scale)
+            else:
+                layer.register_parameter("input_scale", None)
 
     def process_weights_after_loading(self, layer: Module) -> None:
         # If checkpoint not serialized fp8, quantize the weights.
@@ -202,9 +208,23 @@ class Fp8LinearMethod(LinearMethodBase):
             # requantize the logical shards as a single weight.
             else:
                 # Dequant -> Quant with max scale so we can run per tensor.
+                weight = layer.weight
+                weight_scale = layer.weight_scale
+
+                # If rocm, use float8_e4m3fnuz.
+                if is_hip():
+                    weight, weight_scale, input_scale = \
+                        normalize_e4m3fn_to_e4m3fnuz(
+                            weight=weight,
+                            weight_scale=weight_scale,
+                            input_scale=layer.input_scale)
+                    if input_scale is not None:
+                        layer.input_scale = Parameter(input_scale,
+                                                      requires_grad=False)
+
                 weight_scale, weight = requantize_with_max_scale(
-                    weight=layer.weight,
+                    weight=weight,
-                    weight_scale=layer.weight_scale,
+                    weight_scale=weight_scale,
                     logical_widths=layer.logical_widths,
                 )
 
@@ -214,8 +234,6 @@ class Fp8LinearMethod(LinearMethodBase):
             if self.quant_config.activation_scheme == "static":
                 layer.input_scale = Parameter(layer.input_scale.max(),
                                               requires_grad=False)
-            else:
-                layer.input_scale = None
 
         if self.use_marlin:
             prepare_fp8_layer_for_marlin(layer)
@@ -346,10 +364,12 @@ class Fp8MoEMethod(FusedMoEMethodBase):
 
         # If checkpoint is fp16, quantize in place.
         if not self.quant_config.is_checkpoint_fp8_serialized:
+            # If rocm, use float8_e4m3fnuz as dtype
+            fp8_dtype = torch.float8_e4m3fnuz \
+                        if is_hip() else torch.float8_e4m3fn
             w13_weight = torch.empty_like(layer.w13_weight.data,
-                                          dtype=torch.float8_e4m3fn)
+                                          dtype=fp8_dtype)
-            w2_weight = torch.empty_like(layer.w2_weight.data,
+            w2_weight = torch.empty_like(layer.w2_weight.data, dtype=fp8_dtype)
-                                         dtype=torch.float8_e4m3fn)
 
             # Re-initialize w13_scale because we directly quantize
             # merged w13 weights and generate a single scaling factor.
@@ -393,6 +413,32 @@ class Fp8MoEMethod(FusedMoEMethodBase):
                     layer.w13_input_scale.max(), requires_grad=False)
                 layer.w2_input_scale = torch.nn.Parameter(
                     layer.w2_input_scale.max(), requires_grad=False)
+            # If rocm, normalize the weights and scales to e4m3fnuz
+            if is_hip():
+                # Normalize the weights and scales
+                w13_weight, w13_weight_scale, w13_input_scale = \
+                    normalize_e4m3fn_to_e4m3fnuz(
+                        layer.w13_weight, layer.w13_weight_scale,
+                        layer.w13_input_scale)
+                w2_weight, w2_weight_scale, w2_input_scale = \
+                    normalize_e4m3fn_to_e4m3fnuz(
+                        layer.w2_weight, layer.w2_weight_scale,
+                        layer.w2_input_scale)
+                # Reset the parameter
+                layer.w13_weight = torch.nn.Parameter(w13_weight,
+                                                      requires_grad=False)
+                layer.w13_weight_scale = torch.nn.Parameter(
+                    w13_weight_scale, requires_grad=False)
+                if w13_input_scale is not None:
+                    layer.w13_input_scale = torch.nn.Parameter(
+                        w13_input_scale, requires_grad=False)
+                layer.w2_weight = torch.nn.Parameter(w2_weight,
+                                                     requires_grad=False)
+                layer.w2_weight_scale = torch.nn.Parameter(w2_weight_scale,
+                                                           requires_grad=False)
+                if w2_input_scale is not None:
+                    layer.w2_input_scale = torch.nn.Parameter(
+                        w2_input_scale, requires_grad=False)
 
             # Fp8 moe kernel needs single weight scale for w13 per expert.
             # We take the max then dequant and requant each expert.

vllm/model_executor/layers/quantization/utils/w8a8_utils.py

@@ -6,9 +6,19 @@ from torch.nn import Parameter
 from vllm import _custom_ops as ops
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
+from vllm.utils import is_hip
+
+# scaled_mm in pytorch on rocm has a bug that requires always
+# providing scaling factor for result. This value is created
+# as global value to avoid multiple tensor allocations, and
+# can be removed once pytorch fixes the bug.
+TORCH_SCALED_MM_SCALE_RESULT = torch.ones(1).cuda() if is_hip() else None
 
 
 def cutlass_fp8_supported() -> bool:
+    # cutlass is not supported on Rocm
+    if is_hip():
+        return False
     capability = current_platform.get_device_capability()
     capability = capability[0] * 10 + capability[1]
 
@@ -147,13 +157,19 @@ def apply_fp8_linear(
 
         if per_tensor_weights and per_tensor_activations:
             # Fused GEMM_DQ
-            output, _ = torch._scaled_mm(qinput,
+            output = torch._scaled_mm(
+                qinput,
                 weight,
                 out_dtype=input.dtype,
                 scale_a=x_scale,
                 scale_b=weight_scale,
+                scale_result=TORCH_SCALED_MM_SCALE_RESULT,
                 bias=bias)
+            # Since in torch 2.5, scaled_mm only returns single value
+            # This should be removed when vllm-nvidia also moves to 2.5
+            if is_hip():
                 return torch.narrow(output, 0, 0, input.shape[0])
+            return torch.narrow(output[0], 0, 0, input.shape[0])
 
         else:
             # Fallback for channelwise case, where we use unfused DQ
@@ -207,3 +223,27 @@ def apply_int8_linear(
                                  scale_b=weight_scale,
                                  out_dtype=input.dtype,
                                  bias=bias)
+
+
+def normalize_e4m3fn_to_e4m3fnuz(
+    weight: torch.Tensor,
+    weight_scale: torch.Tensor,
+    input_scale: Optional[torch.Tensor] = None
+) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+    assert weight.dtype == torch.float8_e4m3fn
+    # The bits pattern 10000000(-128) represents zero in e4m3fn
+    # but NaN in e4m3fnuz. So here we set it to 0.
+    # https://onnx.ai/onnx/technical/float8.html
+    weight_as_int8 = weight.view(torch.int8)
+    ROCM_FP8_NAN_AS_INT = -128
+    weight_as_int8[weight_as_int8 == ROCM_FP8_NAN_AS_INT] = 0
+    weight = weight_as_int8.view(torch.float8_e4m3fnuz)
+
+    # For the same bits representation, e4m3fnuz value is half of
+    # the e4m3fn value, so we should double the scaling factor to
+    # get the same dequantized value.
+    # https://onnx.ai/onnx/technical/float8.html
+    weight_scale = weight_scale * 2.0
+    if input_scale is not None:
+        input_scale = input_scale * 2.0
+    return weight, weight_scale, input_scale