sglang 0.4.4.post4__py3-none-any.whl → 0.4.5.post1__py3-none-any.whl

sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py

@@ -77,6 +77,7 @@ class CompressedTensorsConfig(QuantizationConfig):
         sparsity_ignore_list: List[str],
         kv_cache_scheme: Optional[Dict[str, Any]] = None,
         config: Optional[Dict[str, Any]] = None,
+        packed_modules_mapping: Dict[str, List[str]] = {},
     ):
         super().__init__()
         self.ignore = ignore
@@ -87,6 +88,7 @@ class CompressedTensorsConfig(QuantizationConfig):
         self.sparsity_scheme_map = sparsity_scheme_map
         self.sparsity_ignore_list = sparsity_ignore_list
         self.config = config
+        self.packed_modules_mapping = packed_modules_mapping
 
     def get_linear_method(self) -> "CompressedTensorsLinearMethod":
         return CompressedTensorsLinearMethod(self)
@@ -136,6 +138,7 @@ class CompressedTensorsConfig(QuantizationConfig):
         sparsity_scheme_map, sparsity_ignore_list = cls._parse_sparsity_config(
             config=config
         )
+        packed_modules_mapping = config.get("packed_modules_mapping", {})
 
         return cls(
             target_scheme_map=target_scheme_map,
@@ -144,6 +147,7 @@ class CompressedTensorsConfig(QuantizationConfig):
             sparsity_scheme_map=sparsity_scheme_map,
             sparsity_ignore_list=sparsity_ignore_list,
             config=config,
+            packed_modules_mapping=packed_modules_mapping,
         )
 
     @classmethod

sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -103,16 +103,6 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
             "input_activations"
         )
 
-        if not (
-            self.weight_quant.strategy == QuantizationStrategy.TENSOR
-            and self.input_quant.strategy == QuantizationStrategy.TENSOR
-        ):
-            raise ValueError(
-                "For FP8 Fused MoE layers, only per-tensor scales "
-                "for weights and activations are supported. Found "
-                f"{self.weight_quant}, {self.input_quant}"
-            )
-
         self.static_input_scales = not self.input_quant.dynamic
 
     def create_weights(
@@ -154,27 +144,50 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
         set_weight_attrs(w2_weight, extra_weight_attrs)
 
         # WEIGHT_SCALES
-        # Allocate 2 scales for w1 and w3 respectively.
-        # They will be combined to a single scale after weight loading.
-        w13_weight_scale = torch.nn.Parameter(
-            torch.ones(num_experts, 2, dtype=torch.float32), requires_grad=False
-        )
-        layer.register_parameter("w13_weight_scale", w13_weight_scale)
+        # per-tensor quantization
+        if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
+            # Allocate 2 scales for w1 and w3 respectively.
+            # They will be combined to a single scale after weight loading.
+            w13_weight_scale = torch.nn.Parameter(
+                torch.ones(num_experts, 2, dtype=torch.float32), requires_grad=False
+            )
+            w2_weight_scale = torch.nn.Parameter(
+                torch.ones(num_experts, dtype=torch.float32), requires_grad=False
+            )
+            weight_quant_method = FusedMoeWeightScaleSupported.TENSOR.value
+        elif self.weight_quant.strategy == QuantizationStrategy.CHANNEL:
+            w13_weight_scale = torch.nn.Parameter(
+                torch.ones(
+                    num_experts,
+                    2 * intermediate_size_per_partition,
+                    1,
+                    dtype=torch.float32,
+                ),
+                requires_grad=False,
+            )
+            w2_weight_scale = torch.nn.Parameter(
+                torch.ones(num_experts, hidden_size, 1, dtype=torch.float32),
+                requires_grad=False,
+            )
+            weight_quant_method = FusedMoeWeightScaleSupported.CHANNEL.value
+        else:
+            raise ValueError(
+                f"Unsupported weight quantization strategy: {self.weight_quant.strategy}"
+            )
 
-        w2_weight_scale = torch.nn.Parameter(
-            torch.ones(num_experts, dtype=torch.float32), requires_grad=False
-        )
+        layer.register_parameter("w13_weight_scale", w13_weight_scale)
         layer.register_parameter("w2_weight_scale", w2_weight_scale)
         # Add the quantization method used (per tensor/grouped/channel)
         # to ensure the weight scales are loaded in properly
-        extra_weight_attrs.update(
-            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
-        )
+        extra_weight_attrs.update({"quant_method": weight_quant_method})
         set_weight_attrs(w13_weight_scale, extra_weight_attrs)
         set_weight_attrs(w2_weight_scale, extra_weight_attrs)
 
         # INPUT_SCALES
         if self.static_input_scales:
+            assert (
+                self.input_quant.strategy == QuantizationStrategy.TENSOR
+            ), "Only per-tensor quantization is supported for static input scales"
             w13_input_scale = torch.nn.Parameter(
                 torch.ones(num_experts, dtype=torch.float32), requires_grad=False
             )
@@ -241,31 +254,37 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
                 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.
-        assert layer.w13_weight_scale is not None
-        shard_size = layer.intermediate_size_per_partition
-        max_w13_scales = layer.w13_weight_scale.max(dim=1).values
-        for expert_id in range(layer.local_num_experts):
-            start = 0
-            for shard_id in range(2):
-                dq_weight = per_tensor_dequantize(
-                    layer.w13_weight[expert_id][start : start + shard_size, :],
-                    layer.w13_weight_scale[expert_id][shard_id],
-                )
-
-                if _is_cuda:
-                    layer.w13_weight[expert_id][start : start + shard_size, :], _ = (
-                        sgl_scaled_fp8_quant(dq_weight, max_w13_scales[expert_id])
-                    )
-                else:
-                    layer.w13_weight[expert_id][start : start + shard_size, :], _ = (
-                        vllm_ops.scaled_fp8_quant(dq_weight, max_w13_scales[expert_id])
+        if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
+            # Fp8 moe kernel needs single weight scale for w13 per expert.
+            # We take the max then dequant and requant each expert.
+            assert layer.w13_weight_scale is not None
+            shard_size = layer.intermediate_size_per_partition
+            max_w13_scales = layer.w13_weight_scale.max(dim=1).values
+            for expert_id in range(layer.local_num_experts):
+                start = 0
+                for shard_id in range(2):
+                    dq_weight = per_tensor_dequantize(
+                        layer.w13_weight[expert_id][start : start + shard_size, :],
+                        layer.w13_weight_scale[expert_id][shard_id],
                     )
-                start += shard_size
 
-        layer.w13_weight_scale = torch.nn.Parameter(max_w13_scales, requires_grad=False)
+                    if _is_cuda:
+                        (
+                            layer.w13_weight[expert_id][start : start + shard_size, :],
+                            _,
+                        ) = sgl_scaled_fp8_quant(dq_weight, max_w13_scales[expert_id])
+                    else:
+                        (
+                            layer.w13_weight[expert_id][start : start + shard_size, :],
+                            _,
+                        ) = vllm_ops.scaled_fp8_quant(
+                            dq_weight, max_w13_scales[expert_id]
+                        )
+                    start += shard_size
+
+            layer.w13_weight_scale = torch.nn.Parameter(
+                max_w13_scales, requires_grad=False
+            )
 
     def apply(
         self,
@@ -285,6 +304,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
         activation: str = "silu",
         inplace: bool = True,
         no_combine: bool = False,
+        apply_router_weight_on_input: bool = False,
     ) -> torch.Tensor:
         from sglang.srt.layers.moe.fused_moe_triton import fused_experts
         from sglang.srt.layers.moe.topk import select_experts
@@ -310,10 +330,13 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
             inplace=inplace,
             activation=activation,
             use_fp8_w8a8=True,
+            per_channel_quant=self.weight_quant.strategy
+            == QuantizationStrategy.CHANNEL,
             w1_scale=layer.w13_weight_scale,
             w2_scale=layer.w2_weight_scale,
             a1_scale=layer.w13_input_scale,
             a2_scale=layer.w2_input_scale,
+            apply_router_weight_on_input=apply_router_weight_on_input,
         )
 
 

sglang/srt/layers/quantization/fp8.py

@@ -71,7 +71,8 @@ ACTIVATION_SCHEMES = ["static", "dynamic"]
 _is_hip = is_hip()
 
 if _is_hip:
-    from aiter.fused_moe_bf16_asm import asm_moe
+    from aiter import ActivationType
+    from aiter.fused_moe_bf16_asm import asm_moe, ck_moe_2stages, ck_moe_2stages_win4
     from aiter.ops.shuffle import shuffle_weight
 
 _is_cuda = is_cuda()
@@ -487,7 +488,7 @@ class Fp8MoEMethod:
 
         if self.quant_config.is_checkpoint_fp8_serialized:
             params_dtype = (
-                torch.int32
+                torch.uint32
                 if get_bool_env_var("USE_INT4_WEIGHT")
                 else torch.float8_e4m3fn
             )
@@ -822,12 +823,14 @@ class Fp8MoEMethod:
         # INT4-FP8 (INT4 MoE Weight, FP8 Compute)
         # Weight Permutation
         layer.w13_weight = torch.nn.Parameter(
-            permute_weight(layer.w13_weight.data),
+            # permute_weight(layer.w13_weight.data),
+            shuffle_weight(layer.w13_weight.data, (16, 16)),
             requires_grad=False,
         )
         torch.cuda.empty_cache()
         layer.w2_weight = torch.nn.Parameter(
-            permute_weight(layer.w2_weight.data),
+            # permute_weight(layer.w2_weight.data),
+            shuffle_weight(layer.w2_weight.data, (16, 16)),
             requires_grad=False,
         )
         torch.cuda.empty_cache()
@@ -860,19 +863,21 @@ class Fp8MoEMethod:
             layer.w13_weight_scale1[expert_id] *= max_w13_scales[expert_id]
             layer.w2_weight_scale1[expert_id] *= layer.w2_weight_scale[expert_id]
 
-    def process_weights_hip_scale_padding(self, layer: Module, padding_size: int):
+    def process_weights_hip_scale_padding(self, layer: Module):
         from sglang.srt.layers.moe.fused_moe_triton.fused_moe import (
             padding_size,  # Avoid circular import
         )
 
         if get_bool_env_var("CK_MOE"):
             layer.w13_weight = torch.nn.Parameter(
-                permute_weight(layer.w13_weight.data),
+                # permute_weight(layer.w13_weight.data),
+                shuffle_weight(layer.w13_weight.data, (16, 16)),
                 requires_grad=False,
             )
             torch.cuda.empty_cache()
             layer.w2_weight = torch.nn.Parameter(
-                permute_weight(layer.w2_weight.data),
+                # permute_weight(layer.w2_weight.data),
+                shuffle_weight(layer.w2_weight.data, (16, 16)),
                 requires_grad=False,
             )
             torch.cuda.empty_cache()
@@ -905,6 +910,7 @@ class Fp8MoEMethod:
         custom_routing_function: Optional[Callable] = None,
         correction_bias: Optional[torch.Tensor] = None,
         activation: str = "silu",
+        apply_router_weight_on_input: bool = False,
         inplace: bool = True,
         no_combine: bool = False,
     ) -> torch.Tensor:
@@ -927,7 +933,7 @@ class Fp8MoEMethod:
         if _is_hip and get_bool_env_var("USE_INT4_WEIGHT"):
             # TODO: add triton kernel and add check get_bool_env_var("CK_MOE")
             assert not no_combine, f"{no_combine=} is not supported."
-            return asm_moe(
+            return ck_moe_2stages_win4(
                 x,
                 layer.w13_weight,
                 layer.w2_weight,
@@ -935,15 +941,17 @@ class Fp8MoEMethod:
                 topk_ids,
                 layer.w13_weight_scale1,
                 layer.w2_weight_scale1,
-                activation=activation,
+                activation=(
+                    ActivationType.Silu if activation == "silu" else ActivationType.Gelu
+                ),
             )
         if _is_hip and get_bool_env_var("CK_MOE"):
-            # TODO(CK_MOE): FP8 or FP8 block_quant only supports 'silu' for the time-being.
-            assert (
-                activation == "silu"
-            ), f"CK_MOE: FP8 and/or FP8 bloack_quant {activation=} will be supported later, unset CK_MOE"
             assert not no_combine, f"{no_combine=} is not supported."
             if self.block_quant:
+                # TODO(CK_MOE): FP8 block_quant only supports 'silu' for the time-being.
+                assert (
+                    activation == "silu"
+                ), f"CK_MOE: FP8 bloack_quant {activation=} will be supported later, unset CK_MOE"
                 return asm_moe(
                     x,
                     layer.w13_weight,
@@ -956,7 +964,7 @@ class Fp8MoEMethod:
                     expert_mask=None,
                 )
             else:
-                return asm_moe(
+                return ck_moe_2stages(
                     x,
                     layer.w13_weight,
                     layer.w2_weight,
@@ -964,6 +972,11 @@ class Fp8MoEMethod:
                     topk_ids,
                     layer.w13_weight_scale1,
                     layer.w2_weight_scale1,
+                    activation=(
+                        ActivationType.Silu
+                        if activation == "silu"
+                        else ActivationType.Gelu
+                    ),
                 )
         else:
             # Expert fusion with FP8 quantization
@@ -975,6 +988,7 @@ class Fp8MoEMethod:
                 topk_ids=topk_ids,
                 inplace=inplace and not no_combine,
                 activation=activation,
+                apply_router_weight_on_input=apply_router_weight_on_input,
                 use_fp8_w8a8=True,
                 w1_scale=(
                     layer.w13_weight_scale_inv

sglang/srt/layers/quantization/fp8_kernel.py

@@ -16,6 +16,7 @@ import functools
 import json
 import logging
 import os
+from contextlib import contextmanager
 from typing import Any, Dict, List, Optional, Tuple
 
 import torch
@@ -40,11 +41,13 @@ fp8_type_ = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
 
 _is_cuda = is_cuda()
 if _is_cuda:
-    import deep_gemm  # `pip install "sgl-kernel>=0.0.4.post3"`
+    import deep_gemm
     from sgl_kernel import sgl_per_token_group_quant_fp8, sgl_per_token_quant_fp8
 
     sm_version = get_device_sm()
-    if sm_version >= 90 and get_bool_env_var("SGL_ENABLE_JIT_DEEPGEMM", default="true"):
+    if sm_version == 90 and get_bool_env_var(
+        "SGL_ENABLE_JIT_DEEPGEMM", default="false"
+    ):
         _enable_jit_deepgemm = True
 
 
@@ -59,7 +62,10 @@ if supports_custom_op():
         Bs: torch.Tensor,
         C: torch.Tensor,
     ) -> None:
-        deep_gemm.gemm_fp8_fp8_bf16_nt((A, As), (B, Bs), C)
+        M, K = A.shape
+        N, _ = B.shape
+        with _log_jit_build(M, N, K):
+            deep_gemm.gemm_fp8_fp8_bf16_nt((A, As), (B, Bs), C)
 
     def deep_gemm_fp8_fp8_bf16_nt_fake(
         A: torch.Tensor,
@@ -708,6 +714,25 @@ def get_w8a8_block_fp8_configs(
     return None
 
 
+@contextmanager
+def _log_jit_build(M: int, N: int, K: int):
+    from deep_gemm.jit.runtime import RuntimeCache
+
+    origin_func = RuntimeCache.__getitem__
+
+    def __patched_func(self, *args, **kwargs):
+        ret = origin_func(self, *args, **kwargs)
+        if ret is None:
+            logger.warning(
+                f"DeepGEMM JIT code generation <gemm_fp8_fp8_bf16_nt>: M={M}, N={N}, K={K}. Please wait."
+            )
+        return ret
+
+    RuntimeCache.__getitem__ = __patched_func
+    yield
+    RuntimeCache.__getitem__ = origin_func
+
+
 def w8a8_block_fp8_matmul(
     A: torch.Tensor,
     B: torch.Tensor,
@@ -782,7 +807,8 @@ def w8a8_block_fp8_matmul(
         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)
+            with _log_jit_build(M, N, K):
+                deep_gemm.gemm_fp8_fp8_bf16_nt((A, As), (B, Bs), C)
     else:
         kernel = (
             _w8a8_block_fp8_matmul_unrolledx4
@@ -815,3 +841,103 @@ def w8a8_block_fp8_matmul(
         )
 
     return C
+
+
+@triton.jit
+def _per_tensor_quant_mla_fp8_stage1(
+    x_ptr,
+    x_s_ptr,
+    head_size,
+    x_stride_h,
+    x_stride_s,
+    eps,
+    fp8_max,
+    BLOCK_SIZE: tl.constexpr,
+):
+    seq_id = tl.program_id(0)
+    head_id = tl.program_id(1)
+    offset = tl.arange(0, BLOCK_SIZE)
+    mask = offset < head_size
+
+    x_ptr += head_id * x_stride_h + seq_id * x_stride_s
+    x = tl.load(x_ptr + offset, mask=mask, other=0.0).to(tl.float32)
+    _absmax = tl.maximum(tl.max(tl.abs(x)), eps)
+
+    tl.atomic_max(x_s_ptr, _absmax / fp8_max)
+
+
+@triton.jit
+def _per_tensor_quant_mla_fp8_stage2(
+    x_ptr,
+    x_s_ptr,
+    x_q_ptr,
+    num_seq,
+    head_size,
+    x_stride_h,
+    x_stride_s,
+    fp8_min,
+    fp8_max,
+    BLOCK_SIZE: tl.constexpr,
+):
+    seq_id = tl.program_id(0)
+    head_id = tl.program_id(1)
+    offset = tl.arange(0, BLOCK_SIZE)
+    mask = offset < head_size
+
+    x_s = tl.load(x_s_ptr)
+    x_s_inv = 1.0 / x_s
+
+    x_ptr += head_id * x_stride_h + seq_id * x_stride_s
+    x_q_ptr += head_id * num_seq * head_size + seq_id * head_size
+
+    x = tl.load(x_ptr + offset, mask=mask, other=0.0).to(tl.float32)
+    x_q = tl.clamp(x * x_s_inv, fp8_min, fp8_max).to(x_q_ptr.dtype.element_ty)
+    tl.store(x_q_ptr + offset, x_q, mask=mask)
+
+
+def per_tensor_quant_mla_fp8(
+    x: torch.Tensor, eps: float = 1e-12, dtype: torch.dtype = torch.float8_e4m3fn
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    This function quantizes input values to float8 values with tensor-wise quantization
+    and specialized for mla absorbed case.
+    """
+    assert x.dim() == 3, "`x` is not a 3d-tensor"
+
+    finfo = torch.finfo(dtype)
+    fp8_max = finfo.max
+    if _is_hip:
+        dtype = torch.float8_e4m3fnuz
+        fp8_max = 224.0
+
+    x_q = x.new_empty(x.size(), dtype=dtype)
+    x_s = torch.zeros((1,), dtype=torch.float32, device=x.device)
+
+    num_head, num_seq, head_size = x.shape
+    BLOCK_SIZE = triton.next_power_of_2(head_size)
+    grid = (num_seq, num_head)
+
+    _per_tensor_quant_mla_fp8_stage1[grid](
+        x,
+        x_s,
+        head_size,
+        x.stride(0),
+        x.stride(1),
+        eps,
+        fp8_max,
+        BLOCK_SIZE,
+    )
+    _per_tensor_quant_mla_fp8_stage2[grid](
+        x,
+        x_s,
+        x_q,
+        num_seq,
+        head_size,
+        x.stride(0),
+        x.stride(1),
+        -fp8_max,
+        fp8_max,
+        BLOCK_SIZE,
+    )
+
+    return x_q, x_s

sglang/srt/layers/quantization/fp8_utils.py

@@ -168,12 +168,13 @@ def input_to_float8(
     """This function quantizes input values to float8 values with tensor-wise quantization."""
     finfo = torch.finfo(dtype)
     min_val, max_val = x.aminmax()
-    amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12)
+    amax = torch.maximum(min_val.abs(), max_val.abs()).float().clamp(min=1e-12)
     fp8_max = finfo.max
     if _is_hip:
+        dtype = torch.float8_e4m3fnuz
         fp8_max = 224.0
     scale = fp8_max / amax
-    x_scl_sat = (x * scale).clamp(min=-fp8_max, max=fp8_max)
+    x_scl_sat = (x.float() * scale).clamp(min=-fp8_max, max=fp8_max)
     return x_scl_sat.to(dtype).contiguous(), scale.float().reciprocal()
 
 
@@ -212,7 +213,24 @@ def block_quant_to_tensor_quant(
         for j in range(n_tiles):
             x_dq_block_tiles[j][i][:, :] = x_dq_block_tiles[j][i] * x_s[j][i]
 
-    x_q_tensor, scale = input_to_float8(x_dq_block, dtype=x_q_block.dtype)
+    x_q_tensor, scale = (
+        sgl_scaled_fp8_quant(x_dq_block)
+        if _is_cuda
+        else input_to_float8(x_dq_block, dtype=x_q_block.dtype)
+    )
+    return x_q_tensor, scale
+
+
+def channel_quant_to_tensor_quant(
+    x_q_channel: torch.Tensor,
+    x_s: torch.Tensor,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    x_dq_channel = x_q_channel.to(torch.float32) * x_s
+    x_q_tensor, scale = (
+        sgl_scaled_fp8_quant(x_dq_channel)
+        if _is_cuda
+        else input_to_float8(x_dq_channel, dtype=x_q_channel.dtype)
+    )
     return x_q_tensor, scale
 
 
@@ -242,9 +260,19 @@ def apply_fp8_linear(
         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]
-            )
+            # TODO(kkhuang): temporarily enforce per-tensor activation scaling if weight is per-tensor scaling
+            # final solution should be: 1. add support to per-tensor activation scaling.
+            # 2. solve the torch.compile error from weight_scale.numel() == 1 and x_scale.numel() > 1 (below line#308)
+            if _is_hip and weight_scale.numel() == 1:
+                qinput, x_scale = ops.scaled_fp8_quant(
+                    input_2d,
+                    input_scale,
+                    use_per_token_if_dynamic=use_per_token_if_dynamic,
+                )
+            else:
+                qinput, x_scale = per_token_group_quant_fp8(
+                    input_2d, group_size=input_2d.shape[1]
+                )
 
     if cutlass_fp8_supported:
         try: