(feat) fp8 fnuz support for rocm

server/text_generation_server/layers/fp8.py

@@ -1,7 +1,7 @@
 import torch
 
 from dataclasses import dataclass
-from typing import Optional, Union, List
+from typing import Optional, Tuple, Union, List
 from loguru import logger
 
 from text_generation_server.utils.import_utils import SYSTEM
@@ -51,8 +51,32 @@ def get_fp8_linear() -> torch.nn.Module:
     return Fp8Linear
 
 
+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
+
+
 def fp8_quantize(
-    weight, scale_upper_bound=None, qdtype=torch.float8_e4m3fn, scalar=False
+    weight, scale=None, scale_upper_bound=None, qdtype=torch.float8_e4m3fn, scalar=False
 ):
     if FBGEMM_DYN_AVAILABLE and not scalar:
         qweight, scale = torch.ops.fbgemm.quantize_fp8_per_row(
@@ -62,8 +86,11 @@ def fp8_quantize(
 
     # weight, scale = quant_weights(weight, torch.int8, False)
     finfo = torch.finfo(qdtype)
-    # Calculate the scale as dtype max divided by absmax
-    scale = finfo.max / weight.abs().max().clamp(min=1e-12, max=scale_upper_bound)
+
+    if scale is None:
+        # Calculate the scale as dtype max divided by absmax
+        scale = finfo.max / weight.abs().max().clamp(min=1e-12, max=scale_upper_bound)
+
     # scale and clamp the tensor to bring it to
     # the representative range of float8 data type
     # (as default cast is unsaturated)
@@ -72,6 +99,10 @@ def fp8_quantize(
     # as both required as inputs to torch._scaled_mm
     qweight = qweight.to(qdtype)
     scale = scale.float().reciprocal()
+
+    if SYSTEM == "rocm":
+        qweight, scale, _ = normalize_e4m3fn_to_e4m3fnuz(qweight, scale)
+
     return qweight, scale
 
 
@@ -92,9 +123,17 @@ class HybridFP8UnquantLoader(WeightsLoader):
                 .reshape(-1)
                 .expand(w.shape[0])
             )
+            try:
+                input_scale = weights.get_tensor(
+                    f"{prefix}.input_scale", to_dtype=False
+                ).reshape(-1)
+            except Exception:
+                input_scale = None
+
             return Fp8Weight(
                 weight=w,
                 weight_scale=scale,
+                input_scale=input_scale,
                 activation_scale_ub=self.activation_scale_ub,
                 dtype=weights.dtype,
             )
@@ -124,10 +163,25 @@ class HybridFP8UnquantLoader(WeightsLoader):
                     to_dtype=False,
                 )
             scale = scale.reshape(-1).expand(w.shape[0])
+            try:
+                input_scale = weights.get_tensor(
+                    f"{prefix}.input_scale", to_dtype=False
+                )
+                if input_scale.numel() > 1:
+                    input_scale = weights.get_packed_sharded(
+                        f"{prefix}.input_scale",
+                        dim=0,
+                        block_sizes=block_sizes,
+                        to_dtype=False,
+                    )
+                input_scale = input_scale.reshape(-1).max()
+            except Exception:
+                input_scale = None
 
             return Fp8Weight(
                 weight=w,
                 weight_scale=scale,
+                input_scale=input_scale,
                 activation_scale_ub=self.activation_scale_ub,
                 dtype=weights.dtype,
             )
@@ -153,10 +207,19 @@ class HybridFP8UnquantLoader(WeightsLoader):
                 for p, shape in zip(prefixes, shapes)
             ]
             scale = torch.cat(scale, dim=0).reshape(-1)
+            try:
+                input_scale = [
+                    _load_scalar_or_matrix_scale(weights, f"{p}.input_scale", shape)
+                    for p, shape in zip(prefixes, shapes)
+                ]
+                input_scale = torch.cat(input_scale, dim=0).reshape(-1).max()
+            except Exception:
+                input_scale = None
 
             return Fp8Weight(
                 weight=w,
                 weight_scale=scale,
+                input_scale=input_scale,
                 activation_scale_ub=self.activation_scale_ub,
                 dtype=weights.dtype,
             )
@@ -174,9 +237,17 @@ class HybridFP8UnquantLoader(WeightsLoader):
                 .reshape(-1)
                 .expand(w.shape[0])
             )
+            try:
+                input_scale = weights.get_tensor(
+                    f"{prefix}.input_scale", to_dtype=False
+                ).reshape(-1)
+            except Exception:
+                input_scale = None
+
             return Fp8Weight(
                 weight=w,
                 weight_scale=scale,
+                input_scale=input_scale,
                 activation_scale_ub=self.activation_scale_ub,
                 dtype=weights.dtype,
             )
@@ -191,6 +262,7 @@ class Fp8Weight(Weight):
     weight: torch.Tensor
     dtype: torch.dtype
     weight_scale: Optional[torch.Tensor] = None
+    input_scale: Optional[torch.Tensor] = None
     activation_scale_ub: Optional[float] = None
 
     def get_linear(self, bias: torch.Tensor):
@@ -200,15 +272,23 @@ class Fp8Weight(Weight):
         # memory. Can be non-contiguous when we e.g. expand from scalars.
         self.weight_scale = self.weight_scale.contiguous()
         return get_fp8_linear().from_fp8(
-            self.weight, self.weight_scale, self.activation_scale_ub, bias, self.dtype
+            self.weight,
+            self.weight_scale,
+            self.input_scale,
+            self.activation_scale_ub,
+            bias,
+            self.dtype,
         )
 
 
 class Fp8Linear(torch.nn.Module):
+    _device_identity_cache = {}
+
     def __init__(
         self,
         qweight,
         scale,
+        input_scale,
         scale_upper_bound,
         bias,
         dtype,
@@ -217,17 +297,29 @@ class Fp8Linear(torch.nn.Module):
         if FBGEMM_MM_AVAILABLE:
             log_once(logger.info, "Using FBGEMM fp8 optimized kernels")
 
+        if SYSTEM == "rocm":
+            qweight, scale, _ = normalize_e4m3fn_to_e4m3fnuz(
+                weight=qweight, weight_scale=scale
+            )
+
         self.dtype = dtype
         self.qweight = qweight
-        self.scale = scale
-        self.scale_upper_bound = (
-            torch.tensor(
-                [scale_upper_bound], dtype=torch.float32, device=qweight.device
-            )
-            if scale_upper_bound is not None
-            else None
+        self.scale = scale.float()
+        self.input_scale = (
+            input_scale.float().reciprocal() if input_scale is not None else None
         )
 
+        if FBGEMM_MM_AVAILABLE:
+            self.scale_upper_bound = (
+                torch.tensor(
+                    [scale_upper_bound], dtype=torch.float32, device=qweight.device
+                )
+                if scale_upper_bound is not None
+                else None
+            )
+        else:
+            self.scale_upper_bound = scale_upper_bound
+
         self.bias = bias if bias is not None else None
 
     @classmethod
@@ -238,18 +330,27 @@ class Fp8Linear(torch.nn.Module):
         )
 
     @classmethod
-    def from_fp8(cls, weight, scale, input_scale, bias, dtype):
+    def from_fp8(cls, weight, scale, input_scale, scale_upper_bound, bias, dtype):
         if FBGEMM_DYN_AVAILABLE:
             # fbgemm needs float32 scales.
             scale = scale.float()
         return cls(
             qweight=weight,
             scale=scale,
-            scale_upper_bound=input_scale,
+            input_scale=input_scale,
+            scale_upper_bound=scale_upper_bound,
             bias=bias,
             dtype=dtype,
         )
 
+    @classmethod
+    def get_shared_device_identity(cls, device):
+        # Input scaling factors are no longer optional in _scaled_mm starting
+        # from pytorch 2.5. Allocating a dummy tensor to pass as input_scale
+        if device not in cls._device_identity_cache:
+            cls._device_identity_cache[device] = torch.ones(1, device=device)
+        return cls._device_identity_cache[device]
+
     def forward(self, input: torch.Tensor) -> torch.Tensor:
         if FBGEMM_MM_AVAILABLE:
             qinput, scale = fp8_quantize(
@@ -266,15 +367,49 @@ class Fp8Linear(torch.nn.Module):
             )
             return y.to(self.dtype)
 
-        qinput, scale = fp8_quantize(input, scalar=True)
-        output, _ = torch._scaled_mm(
-            qinput,
-            self.qweight.t(),
-            out_dtype=self.dtype,
-            scale_a=scale,
-            scale_b=self.scale,
-            bias=self.bias,
+        qinput, scale = fp8_quantize(
+            input,
+            self.input_scale,
+            scale_upper_bound=self.scale_upper_bound,
+            scalar=True,
         )
+
+        per_tensor_weights = self.scale.numel() == 1
+        per_tensor_activations = scale.numel() == 1
+
+        if SYSTEM != "rocm" or (per_tensor_weights and per_tensor_activations):
+            output = torch._scaled_mm(
+                qinput,
+                self.qweight.t(),
+                out_dtype=self.dtype,
+                scale_a=scale,
+                scale_b=self.scale,
+                bias=self.bias,
+            )
+
+            if type(output) is tuple and len(output) == 2:
+                output = output[0]
+        else:
+            device_identity = None
+            if SYSTEM == "rocm":
+                device_identity = self.get_shared_device_identity(self.qweight.device)
+
+            output = torch._scaled_mm(
+                qinput,
+                self.qweight.t(),
+                scale_a=device_identity,
+                scale_b=device_identity,
+                out_dtype=torch.float32,
+            )
+            if type(output) is tuple and len(output) == 2:
+                output = output[0]
+
+            output = output * scale * self.scale.t()
+            if self.bias is not None:
+                output = output + self.bias
+
+            output = output.to(dtype=self.dtype)
+
         return output
 
 

server/text_generation_server/layers/marlin/fp8.py

@@ -62,7 +62,7 @@ class GPTQMarlinFP8Linear(nn.Module):
         return cls(qweight=qweight, scales=scales.to(dtype), bias=bias)
 
     @classmethod
-    def from_fp8(cls, weight, scale, _input_scale, bias, dtype):
+    def from_fp8(cls, weight, scale, _input_scale, _scale_upper_bound, bias, dtype):
         return cls(qweight=weight, scales=scale.to(dtype), bias=bias)
 
     def forward(self, A: torch.Tensor) -> torch.Tensor:

server/text_generation_server/models/__init__.py

@@ -340,7 +340,7 @@ def get_model(
         if method in {"gptq", "awq", "exl2"}:
             log_master(logger.info, f"Auto selecting quantization method {method}")
             quantize = method
-        elif method == "fbgemm_fp8":
+        elif method == "fbgemm_fp8" or method == "fp8":
             log_master(logger.info, "Auto selecting quantization method fp8")
             quantize = "fp8"
         else: