sglang 0.4.3.post4__py3-none-any.whl → 0.4.4.post1__py3-none-any.whl

sglang/srt/layers/quantization/fp8_kernel.py

@@ -22,17 +22,54 @@ import torch
 import triton
 import triton.language as tl
 
-from sglang.srt.utils import get_device_core_count, get_device_name, is_hip
-
-is_hip_ = is_hip()
-fp8_type_ = torch.float8_e4m3fnuz if is_hip_ else torch.float8_e4m3fn
-
-_is_cuda = torch.cuda.is_available() and torch.version.cuda
+from sglang.srt.utils import (
+    direct_register_custom_op,
+    get_device_core_count,
+    get_device_name,
+    is_cuda,
+    is_hip,
+    supports_custom_op,
+)
+
+_is_hip = is_hip()
+fp8_type_ = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
+
+_is_cuda = is_cuda()
 if _is_cuda:
-    from sgl_kernel import sgl_per_token_group_quant_fp8
+    import deep_gemm  # `pip install "sgl-kernel>=0.0.4.post3"`
+    from sgl_kernel import sgl_per_token_group_quant_fp8, sgl_per_token_quant_fp8
 
 logger = logging.getLogger(__name__)
 
+_enable_jit_deepgemm = int(os.getenv("SGL_ENABLE_JIT_DEEPGEMM", "0"))
+
+if supports_custom_op():
+
+    def deep_gemm_fp8_fp8_bf16_nt(
+        A: torch.Tensor,
+        As: torch.Tensor,
+        B: torch.Tensor,
+        Bs: torch.Tensor,
+        C: torch.Tensor,
+    ) -> None:
+        deep_gemm.gemm_fp8_fp8_bf16_nt((A, As), (B, Bs), C)
+
+    def deep_gemm_fp8_fp8_bf16_nt_fake(
+        A: torch.Tensor,
+        As: torch.Tensor,
+        B: torch.Tensor,
+        Bs: torch.Tensor,
+        C: torch.Tensor,
+    ) -> None:
+        return
+
+    direct_register_custom_op(
+        op_name="deep_gemm_fp8_fp8_bf16_nt",
+        op_func=deep_gemm_fp8_fp8_bf16_nt,
+        mutates_args=["C"],
+        fake_impl=deep_gemm_fp8_fp8_bf16_nt_fake,
+    )
+
 
 @triton.jit
 def _per_token_group_quant_fp8(
@@ -70,7 +107,8 @@ def _per_token_group_quant_fp8(
     # Quant
     _absmax = tl.maximum(tl.max(tl.abs(y)), eps)
     y_s = _absmax / fp8_max
-    y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
+    y_s_inv = 1.0 / y_s
+    y_q = tl.clamp(y * y_s_inv, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
 
     tl.store(y_q_ptr + cols, y_q, mask=mask)
     tl.store(y_s_ptr, y_s)
@@ -140,7 +178,7 @@ def per_token_group_quant_fp8(
         x: The input tenosr with ndim >= 2.
         group_size: The group size used for quantization.
         eps: The minimum to avoid dividing zero.
-        dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn` is supported for now.
+        dtype: The dype of output tensor.
 
     Returns:
         Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the scaling factor for quantization.
@@ -153,7 +191,7 @@ def per_token_group_quant_fp8(
     finfo = torch.finfo(dtype)
     fp8_max = finfo.max
 
-    if is_hip_:
+    if _is_hip:
         fp8_max = 224.0
 
     fp8_min = -fp8_max
@@ -241,6 +279,132 @@ def sglang_per_token_group_quant_fp8(
     return x_q, x_s
 
 
+def sglang_per_token_quant_fp8(
+    x: torch.Tensor,
+    dtype: torch.dtype = fp8_type_,
+):
+    assert x.is_contiguous(), "`x` is not contiguous"
+
+    x_q = torch.empty_like(x, device=x.device, dtype=dtype)
+    x_s = torch.empty(
+        x.shape[0],
+        1,
+        device=x.device,
+        dtype=torch.float32,
+    )
+
+    sgl_per_token_quant_fp8(x, x_q, x_s)
+
+    return x_q, x_s
+
+
+@triton.jit
+def _static_quant_fp8(
+    # Pointers to inputs and output
+    y_ptr,
+    y_q_ptr,
+    y_s_ptr,
+    y_s_repeat_ptr,
+    # Stride of input
+    y_stride,
+    # Collums of input
+    N,
+    # Information for float8
+    fp8_min,
+    fp8_max,
+    # Meta-parameters
+    BLOCK: tl.constexpr,
+    REPEAT_SCALE: tl.constexpr,
+):
+    """A Triton-accelerated function to perform quantization using the given scale on a
+    tensor
+
+    This function converts the tensor values into float8 values.
+    """
+    # Map the program id to the row of X and Y it should compute.
+    g_id = tl.program_id(0)
+    y_ptr += g_id * y_stride
+    y_q_ptr += g_id * y_stride
+    if REPEAT_SCALE:
+        y_s_repeat_ptr += g_id
+
+    cols = tl.arange(0, BLOCK)  # N <= BLOCK
+    mask = cols < N
+
+    y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32)
+    y_s = tl.load(y_s_ptr).to(tl.float32)
+    y_s_inv = 1.0 / y_s
+    y_q = tl.clamp(y * y_s_inv, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
+
+    tl.store(y_q_ptr + cols, y_q, mask=mask)
+    if REPEAT_SCALE:
+        tl.store(y_s_repeat_ptr, y_s)
+
+
+def static_quant_fp8(
+    x: torch.Tensor,
+    x_s: torch.Tensor,
+    repeat_scale: bool = False,
+    dtype: torch.dtype = fp8_type_,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Function to perform static quantization using the given scale on an input tensor `x`.
+
+    It converts the tensor values into signed float8 values and returns the
+    quantized tensor along with the scaling factor used for quantization.
+
+    Args:
+        x: The input tenosr with ndim >= 2.
+        x_s: The quantization scale.
+        repeat_scale: Whether to broadcast per-tensor scale to per-channel scale.
+        dtype: The dype of output tensor.
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the scaling factor for quantization.
+    """
+    assert x.is_contiguous(), "`x` is not contiguous"
+    assert x_s.numel() == 1, "only supports per-tensor scale"
+    finfo = torch.finfo(dtype)
+    fp8_max = finfo.max
+
+    if _is_hip:
+        fp8_max = 224.0
+
+    fp8_min = -fp8_max
+
+    x_q = torch.empty_like(x, device=x.device, dtype=dtype)
+    M = x.numel() // x.shape[-1]
+    N = x.shape[-1]
+    if repeat_scale:
+        x_s_repeat = torch.empty(
+            (M, 1),
+            device=x.device,
+            dtype=torch.float32,
+        )
+    else:
+        x_s_repeat = None
+
+    BLOCK = triton.next_power_of_2(N)
+    # heuristics for number of warps
+    num_warps = min(max(BLOCK // 256, 1), 8)
+    num_stages = 1
+    _static_quant_fp8[(M,)](
+        x,
+        x_q,
+        x_s,
+        x_s_repeat,
+        N,
+        N,
+        fp8_min=fp8_min,
+        fp8_max=fp8_max,
+        BLOCK=BLOCK,
+        REPEAT_SCALE=repeat_scale,
+        num_warps=num_warps,
+        num_stages=num_stages,
+    )
+    x_s = x_s_repeat if repeat_scale else x_s
+    return x_q, x_s
+
+
 @triton.jit
 def _w8a8_block_fp8_matmul(
     # Pointers to inputs and output
@@ -595,34 +759,42 @@ def w8a8_block_fp8_matmul(
     num_workgroups = triton.cdiv(M, config["BLOCK_SIZE_M"]) * triton.cdiv(
         N, config["BLOCK_SIZE_N"]
     )
-    kernel = (
-        _w8a8_block_fp8_matmul_unrolledx4
-        if (is_hip_ == True and num_workgroups <= get_device_core_count())
-        else _w8a8_block_fp8_matmul
-    )
 
-    kernel[grid](
-        A,
-        B,
-        C,
-        As,
-        Bs,
-        M,
-        N,
-        K,
-        block_n,
-        block_k,
-        A.stride(-2),
-        A.stride(-1),
-        B.stride(1),
-        B.stride(0),
-        C.stride(-2),
-        C.stride(-1),
-        As.stride(-2),
-        As.stride(-1),
-        Bs.stride(1),
-        Bs.stride(0),
-        **config,
-    )
+    # deepgemm only support bf16
+    if _is_cuda and C.dtype == torch.bfloat16 and _enable_jit_deepgemm:
+        if supports_custom_op():
+            torch.ops.sglang.deep_gemm_fp8_fp8_bf16_nt(A, As, B, Bs, C)
+        else:
+            deep_gemm.gemm_fp8_fp8_bf16_nt((A, As), (B, Bs), C)
+    else:
+        kernel = (
+            _w8a8_block_fp8_matmul_unrolledx4
+            if (_is_hip == True and num_workgroups <= get_device_core_count())
+            else _w8a8_block_fp8_matmul
+        )
+
+        kernel[grid](
+            A,
+            B,
+            C,
+            As,
+            Bs,
+            M,
+            N,
+            K,
+            block_n,
+            block_k,
+            A.stride(-2),
+            A.stride(-1),
+            B.stride(1),
+            B.stride(0),
+            C.stride(-2),
+            C.stride(-1),
+            As.stride(-2),
+            As.stride(-1),
+            Bs.stride(1),
+            Bs.stride(0),
+            **config,
+        )
 
     return C

sglang/srt/layers/quantization/fp8_utils.py

@@ -1,23 +1,53 @@
-import os
 from typing import List, Optional, Tuple
 
 import torch
 
-from sglang.srt.layers.parameter import RowvLLMParameter, _ColumnvLLMParameter
 from sglang.srt.layers.quantization.fp8_kernel import (
     per_token_group_quant_fp8,
+    static_quant_fp8,
     w8a8_block_fp8_matmul,
 )
-from sglang.srt.utils import get_bool_env_var, is_hip
+from sglang.srt.utils import (
+    get_bool_env_var,
+    get_cuda_version,
+    get_device_capability,
+    is_cuda,
+    is_hip,
+)
+
+use_vllm_cutlass_w8a8_fp8_kernel = get_bool_env_var("USE_VLLM_CUTLASS_W8A8_FP8_KERNEL")
 
-is_hip_ = is_hip()
-if is_hip_ and get_bool_env_var("CK_MOE"):
+_is_hip = is_hip()
+if _is_hip and get_bool_env_var("CK_MOE"):
     from aiter import gemm_a8w8_blockscale
 
-_is_cuda = torch.cuda.is_available() and torch.version.cuda
+_is_cuda = is_cuda()
 if _is_cuda:
     from sgl_kernel import fp8_blockwise_scaled_mm
 
+    from sglang.srt.layers.quantization.fp8_kernel import sglang_per_token_quant_fp8
+
+    if use_vllm_cutlass_w8a8_fp8_kernel:
+        from vllm import _custom_ops as ops
+    else:
+        from sgl_kernel import fp8_scaled_mm
+
+# Input scaling factors are no longer optional in _scaled_mm starting
+# from pytorch 2.5. Allocating a dummy tensor to pass as input_scale
+TORCH_DEVICE_IDENTITY = torch.ones(1, dtype=torch.float32)
+
+
+def cutlass_fp8_supported():
+    if not _is_cuda:
+        return False
+    major, minor = get_device_capability()
+    cuda_version = get_cuda_version()
+    if major >= 9:
+        return cuda_version >= (12, 0)
+    elif major == 8 and minor == 9:
+        return cuda_version >= (12, 4)
+    return False
+
 
 def normalize_e4m3fn_to_e4m3fnuz(
     weight: torch.Tensor,
@@ -44,7 +74,7 @@ def normalize_e4m3fn_to_e4m3fnuz(
 
 
 def cutlass_block_fp8_supported() -> bool:
-    if os.environ.get("SUPPORT_CUTLASS_BLOCK_FP8") is None:
+    if get_bool_env_var("SUPPORT_CUTLASS_BLOCK_FP8"):
         return False
     if _is_cuda:
         major, minor = torch.cuda.get_device_capability()
@@ -81,7 +111,7 @@ def apply_w8a8_block_fp8_linear(
         output = fp8_blockwise_scaled_mm(
             q_input, weight.T, x_scale, weight_scale.T, out_dtype=input.dtype
         )
-    elif is_hip_ and get_bool_env_var("CK_MOE"):
+    elif _is_hip and get_bool_env_var("CK_MOE"):
         q_input, x_scale = per_token_group_quant_fp8(
             input_2d, block_size[1], column_major_scales=False
         )
@@ -112,7 +142,7 @@ def input_to_float8(
     min_val, max_val = x.aminmax()
     amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12)
     fp8_max = finfo.max
-    if is_hip_:
+    if _is_hip:
         fp8_max = 224.0
     scale = fp8_max / amax
     x_scl_sat = (x * scale).clamp(min=-fp8_max, max=fp8_max)
@@ -158,10 +188,121 @@ def block_quant_to_tensor_quant(
     return x_q_tensor, scale
 
 
-class BlockQuantScaleParameter(_ColumnvLLMParameter, RowvLLMParameter):
-    """
-    Parameter class for weight scales loaded for weights with
-    block-wise quantization. Uses both column and row parallelism.
-    """
+def apply_fp8_linear(
+    input: torch.Tensor,
+    weight: torch.Tensor,
+    weight_scale: torch.Tensor,
+    input_scale: Optional[torch.Tensor] = None,
+    input_scale_ub: Optional[torch.Tensor] = None,
+    bias: Optional[torch.Tensor] = None,
+    cutlass_fp8_supported: bool = True,
+    use_per_token_if_dynamic: bool = False,
+) -> torch.Tensor:
+    # View input as 2D matrix for fp8 methods
+    input_2d = input.view(-1, input.shape[-1])
+    output_shape = [*input.shape[:-1], weight.shape[1]]
+
+    # cutlass w8a8 fp8 sgl-kernel only supports per-token scale
+    if input_scale is not None:
+        assert input_scale.numel() == 1
+        # broadcast per-tensor scale to per-token scale when supporting cutlass
+        qinput, x_scale = static_quant_fp8(
+            input_2d, input_scale, repeat_scale=cutlass_fp8_supported
+        )
+    else:
+        # default use per-token quantization if dynamic
+        if _is_cuda:
+            qinput, x_scale = sglang_per_token_quant_fp8(input_2d)
+        else:
+            qinput, x_scale = per_token_group_quant_fp8(
+                input_2d, group_size=input_2d.shape[1]
+            )
+
+    if cutlass_fp8_supported:
+        if use_vllm_cutlass_w8a8_fp8_kernel:
+            # Fall back to vllm cutlass w8a8 fp8 kernel
+            output = ops.cutlass_scaled_mm(
+                qinput,
+                weight,
+                out_dtype=input.dtype,
+                scale_a=x_scale,
+                scale_b=weight_scale,
+                bias=bias,
+            )
+        else:
+            assert (
+                weight_scale.numel() == weight.shape[1]
+            ), "cutlass w8a8 fp8 sgl-kernel only supports per-channel scale"
+            output = fp8_scaled_mm(
+                qinput, weight, x_scale, weight_scale, out_dtype=input.dtype, bias=bias
+            )
+        return output.view(*output_shape)
+
+    # torch.scaled_mm supports per tensor weights + activations only
+    # so fallback to naive if per channel or per token
+    else:
+        per_tensor_weights = weight_scale.numel() == 1
+        per_tensor_activations = x_scale.numel() == 1
+
+        if per_tensor_weights and per_tensor_activations:
+            # Fused GEMM_DQ
+            output = torch._scaled_mm(
+                qinput,
+                weight,
+                out_dtype=input.dtype,
+                scale_a=x_scale,
+                scale_b=weight_scale,
+                bias=bias,
+            )
+            # A fix for discrepancy in scaled_mm which returns tuple
+            # for torch < 2.5 and a single value in torch >= 2.5
+            if type(output) is tuple and len(output) == 2:
+                output = output[0]
+
+            return torch.narrow(output, 0, 0, input_2d.shape[0]).view(*output_shape)
+
+        else:
+            # Fallback for channelwise case, where we use unfused DQ
+            # due to limitations with scaled_mm
+
+            # Symmetric quantized GEMM by definition computes the following:
+            #   C = (s_x * X) (s_w * W) + bias
+            # This is equivalent to dequantizing the weights and activations
+            # before applying a GEMM.
+            #
+            # In order to compute quantized operands, a quantized kernel
+            # will rewrite the above like so:
+            #   C = s_w * s_x * (X * W) + bias
+            #
+            # For the scaled_mm fallback case, we break this down, since it
+            # does not support s_w being a vector.
+
+            # Making sure the dummy tensor is on the same device as the weight
+            global TORCH_DEVICE_IDENTITY
+            if TORCH_DEVICE_IDENTITY.device != weight.device:
+                TORCH_DEVICE_IDENTITY = TORCH_DEVICE_IDENTITY.to(weight.device)
+
+            # GEMM
+            # This computes C = (X * W).
+            # Output in fp32 to allow subsequent ops to happen in-place
+            output = torch._scaled_mm(
+                qinput,
+                weight,
+                scale_a=TORCH_DEVICE_IDENTITY,
+                scale_b=TORCH_DEVICE_IDENTITY,
+                out_dtype=torch.float32,
+            )
+            # A fix for discrepancy in scaled_mm which returns tuple
+            # for torch < 2.5 and a single value in torch >= 2.5
+            if type(output) is tuple and len(output) == 2:
+                output = output[0]
+            # Unpad (undo num_token_padding)
+            output = torch.narrow(output, 0, 0, input_2d.shape[0])
+            x_scale = torch.narrow(x_scale, 0, 0, input_2d.shape[0])
 
-    pass
+            # DQ
+            # C = sw * sx * (X * W) + bias
+            output = output * x_scale * weight_scale.t()
+            if bias is not None:
+                output = output + bias
+            return output.to(dtype=input.dtype).view(*output_shape)

sglang/srt/layers/quantization/modelopt_quant.py

@@ -7,7 +7,7 @@ import torch
 from torch.nn.parameter import Parameter
 from vllm.model_executor.layers.quantization.kv_cache import BaseKVCacheMethod
 from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
-    apply_fp8_linear,
+    convert_to_channelwise,
     cutlass_fp8_supported,
     requantize_with_max_scale,
 )
@@ -19,6 +19,7 @@ from sglang.srt.layers.quantization.base_config import (
     QuantizationConfig,
     QuantizeMethodBase,
 )
+from sglang.srt.layers.quantization.fp8_utils import apply_fp8_linear
 
 # Initialize logger for the module
 logger = logging.getLogger(__name__)
@@ -161,6 +162,9 @@ class ModelOptFp8LinearMethod(LinearMethodBase):
             layer.weight, layer.weight_scale, layer.logical_widths
         )
         layer.weight = Parameter(quantized_weight.t(), requires_grad=False)
+        # cutlass sgl-kernel only supports per-channel scale
+        if self.cutlass_fp8_supported:
+            max_w_scale = convert_to_channelwise(max_w_scale, layer.logical_widths)
         layer.weight_scale = Parameter(max_w_scale, requires_grad=False)
         layer.input_scale = Parameter(layer.input_scale.max(), requires_grad=False)
 

sglang/srt/layers/quantization/w8a8_fp8.py

@@ -0,0 +1,128 @@
+from typing import Any, Dict, List, Optional
+
+import torch
+from torch.nn.parameter import Parameter
+
+from sglang.srt.layers.linear import LinearMethodBase
+from sglang.srt.layers.parameter import ChannelQuantScaleParameter, ModelWeightParameter
+from sglang.srt.layers.quantization.base_config import (
+    QuantizationConfig,
+    QuantizeMethodBase,
+)
+from sglang.srt.layers.quantization.fp8_utils import (
+    apply_fp8_linear,
+    cutlass_fp8_supported,
+    normalize_e4m3fn_to_e4m3fnuz,
+)
+from sglang.srt.utils import is_hip
+
+_is_hip = is_hip()
+
+
+class W8A8Fp8Config(QuantizationConfig):
+    """Config class for W8A8 FP8 Quantization.
+
+    - Weight: static, per-channel, symmetric
+    - Activation: dynamic, per-token, symmetric
+    """
+
+    def __init__(self):
+        pass
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> List[torch.dtype]:
+        return [torch.float16, torch.bfloat16]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 89
+
+    @classmethod
+    def get_name(self) -> str:
+        return "w8a8_fp8"
+
+    @classmethod
+    def get_config_filenames(cls) -> List[str]:
+        return []
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> "W8A8Fp8Config":
+        return cls()
+
+    def get_quant_method(
+        self,
+        layer: torch.nn.Module,
+        prefix: str,
+    ) -> Optional["QuantizeMethodBase"]:
+        from sglang.srt.layers.linear import LinearBase
+
+        if isinstance(layer, LinearBase):
+            return W8A8Fp8LinearMethod(self)
+        return None
+
+    def get_scaled_act_names(self) -> List[str]:
+        return []
+
+
+class W8A8Fp8LinearMethod(LinearMethodBase):
+
+    def __init__(self, quantization_config: W8A8Fp8Config):
+        self.cutlass_fp8_supported = cutlass_fp8_supported()
+        self.quantization_config = quantization_config
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        weight = layer.weight
+        weight_scale = layer.weight_scale.detach()
+        if _is_hip:
+            weight, weight_scale, _ = normalize_e4m3fn_to_e4m3fnuz(
+                weight=weight, weight_scale=weight_scale
+            )
+        layer.weight = Parameter(weight.t(), requires_grad=False)
+        layer.weight_scale = Parameter(weight_scale, requires_grad=False)
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size_per_partition: int,
+        output_partition_sizes: List[int],
+        input_size: int,
+        output_size: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs
+    ):
+
+        weight_loader = extra_weight_attrs.get("weight_loader")
+        self.logical_widths = output_partition_sizes
+
+        weight = ModelWeightParameter(
+            data=torch.empty(
+                sum(output_partition_sizes),
+                input_size_per_partition,
+                dtype=torch.float8_e4m3fn,
+            ),
+            input_dim=1,
+            output_dim=0,
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("weight", weight)
+
+        weight_scale = ChannelQuantScaleParameter(
+            data=torch.empty((sum(output_partition_sizes), 1), dtype=torch.float32),
+            output_dim=0,
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("weight_scale", weight_scale)
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+    ):
+        return apply_fp8_linear(
+            x,
+            layer.weight,
+            layer.weight_scale,
+            bias=bias,
+            cutlass_fp8_supported=self.cutlass_fp8_supported,
+        )