[Feature] DeepSeek V3/R1 INT8 Quantization (channel-wise) (#3888)

Co-authored-by: yych0745 <1398089567@qq.com>
Co-authored-by: sleepcoo <sleepcoo@gmail.com>
Co-authored-by: b0urnee <2769086541@qq.com>

python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py

@@ -15,7 +15,10 @@ from vllm import _custom_ops as ops
 
 from sglang.srt.layers.moe.topk import select_experts
 from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
-from sglang.srt.layers.quantization.int8_kernel import per_token_group_quant_int8
+from sglang.srt.layers.quantization.int8_kernel import (
+    per_token_group_quant_int8,
+    per_token_quant_int8,
+)
 from sglang.srt.utils import (
     direct_register_custom_op,
     get_bool_env_var,
@@ -117,6 +120,7 @@ def fused_moe_kernel(
     - expert_ids: A tensor containing the indices of the expert for each
         block. It determines which expert matrix from B should be used for
         each block in A.
+
     This kernel performs the multiplication of a token by its corresponding
     expert matrix as determined by `expert_ids`. The sorting of
     `sorted_token_ids` by expert index and padding ensures divisibility by
@@ -167,17 +171,38 @@ def fused_moe_kernel(
         )
         b_scale = tl.load(b_scale_ptrs)
 
-    if use_fp8_w8a8 or use_int8_w8a8:
+    if use_fp8_w8a8:
+        # block-wise
         if group_k > 0 and group_n > 0:
             a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
             offs_bsn = offs_bn // group_n
             b_scale_ptrs = (
                 b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn
             )
+        # tensor-wise
         else:
             a_scale = tl.load(a_scale_ptr)
             b_scale = tl.load(b_scale_ptr + off_experts)
 
+    if use_int8_w8a8:
+        # block-wise
+        if group_k > 0 and group_n > 0:
+            a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
+            offs_bsn = offs_bn // group_n
+            b_scale_ptrs = (
+                b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn
+            )
+        # channel-wise
+        else:
+            # Load per-column scale for weights
+            b_scale_ptrs = (
+                b_scale_ptr + off_experts * stride_bse + offs_bn[None, :] * stride_bsn
+            )
+            b_scale = tl.load(b_scale_ptrs)
+            # Load per-token scale for activations
+            a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm
+            a_scale = tl.load(a_scale_ptrs, mask=token_mask, other=0.0)[:, None]
+
     # -----------------------------------------------------------
     # Iterate to compute a block of the C matrix.
     # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
@@ -217,9 +242,13 @@ def fused_moe_kernel(
 
                 accumulator += tl.dot(a, b) * a_scale[:, None] * b_scale[None, :]
             else:
+                # fix out of shared memory issue
+                if use_fp8_w8a8:
                     accumulator = tl.dot(a, b, acc=accumulator)
                 else:
                     accumulator += tl.dot(a, b)
+        else:
+            accumulator += tl.dot(a, b)
         # Advance the ptrs to the next K block.
         a_ptrs += BLOCK_SIZE_K * stride_ak
         b_ptrs += BLOCK_SIZE_K * stride_bk
@@ -497,9 +526,11 @@ def invoke_fused_moe_kernel(
     if use_fp8_w8a8:
         assert B_scale is not None
         if block_shape is None:
+            # activation tensor-wise fp8 quantization, dynamic or static
             padded_size = padding_size
             A, A_scale = ops.scaled_fp8_quant(A, A_scale)
         else:
+            # activation block-wise fp8 quantization
             assert len(block_shape) == 2
             block_n, block_k = block_shape[0], block_shape[1]
             if _is_cuda:
@@ -512,9 +543,10 @@ def invoke_fused_moe_kernel(
     elif use_int8_w8a8:
         assert B_scale is not None
         if block_shape is None:
-            padded_size = padding_size
-            A, A_scale = ops.scaled_int8_quant(A, A_scale)
+            # activation channel-wise int8 quantization
+            A, A_scale = per_token_quant_int8(A)
         else:
+            # activation block-wise int8 quantization
             assert len(block_shape) == 2
             block_n, block_k = block_shape[0], block_shape[1]
             A, A_scale = per_token_group_quant_int8(A, block_k)
@@ -1060,7 +1092,6 @@ def fused_experts_impl(
             use_int8_w8a16=use_int8_w8a16,
             block_shape=block_shape,
         )
-
         if activation == "silu":
             if _is_cuda:
                 silu_and_mul(intermediate_cache1.view(-1, N), intermediate_cache2)

python/sglang/srt/layers/quantization/w8a8_int8.py

-from typing import Any, Dict, List, Optional
+from typing import Any, Callable, Dict, List, Optional
 
 import torch
 
-from sglang.srt.utils import is_cuda_available
+from sglang.srt.utils import is_cuda_available, set_weight_attrs
 
 is_cuda = is_cuda_available()
 if is_cuda:
@@ -10,6 +10,7 @@ if is_cuda:
 
 from torch.nn.parameter import Parameter
 
+from sglang.srt.distributed import get_tensor_model_parallel_world_size
 from sglang.srt.layers.linear import LinearMethodBase
 from sglang.srt.layers.parameter import ChannelQuantScaleParameter, ModelWeightParameter
 from sglang.srt.layers.quantization.base_config import (
@@ -55,9 +56,12 @@ class W8A8Int8Config(QuantizationConfig):
         prefix: str,
     ) -> Optional["QuantizeMethodBase"]:
         from sglang.srt.layers.linear import LinearBase
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 
         if isinstance(layer, LinearBase):
             return W8A8Int8LinearMethod(self)
+        elif isinstance(layer, FusedMoE):
+            return W8A8Int8MoEMethod(self)
         return None
 
     def get_scaled_act_names(self) -> List[str]:
@@ -81,7 +85,7 @@ class W8A8Int8LinearMethod(LinearMethodBase):
         input_size: int,
         output_size: int,
         params_dtype: torch.dtype,
-        **extra_weight_attrs
+        **extra_weight_attrs,
     ):
 
         weight_loader = extra_weight_attrs.get("weight_loader")
@@ -115,3 +119,148 @@ class W8A8Int8LinearMethod(LinearMethodBase):
         return int8_scaled_mm(
             x_q, layer.weight, x_scale, layer.weight_scale, out_dtype=x.dtype, bias=bias
         )
+
+
+class W8A8Int8MoEMethod:
+    """MoE method for INT8.
+    Supports loading INT8 checkpoints with static weight scale and
+    dynamic/static activation scale.
+    Also supports loading quantized FP16/BF16 model checkpoints with dynamic
+    activation scaling. The weight scaling factor will be initialized after
+    the model weights are loaded.
+    Args:
+        quant_config: The quantization config.
+    """
+
+    def __new__(cls, *args, **kwargs):
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoEMethodBase
+
+        if not hasattr(cls, "_initialized"):
+            original_init = cls.__init__
+            new_cls = type(
+                cls.__name__,
+                (FusedMoEMethodBase,),
+                {
+                    "__init__": original_init,
+                    **{k: v for k, v in cls.__dict__.items() if k != "__dict__"},
+                },
+            )
+            obj = super(new_cls, new_cls).__new__(new_cls)
+            obj.__init__(*args, **kwargs)
+            return obj
+        return super().__new__(cls)
+
+    def __init__(self, quant_config):
+        self.quant_config = quant_config
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
+
+        tp_size = get_tensor_model_parallel_world_size()
+
+        # WEIGHTS
+        w13_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts, 2 * intermediate_size, hidden_size, dtype=torch.int8
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight", w13_weight)
+        set_weight_attrs(w13_weight, extra_weight_attrs)
+
+        w2_weight = torch.nn.Parameter(
+            torch.empty(num_experts, hidden_size, intermediate_size, dtype=torch.int8),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        w13_weight_scale = torch.nn.Parameter(
+            torch.ones(num_experts, 2 * intermediate_size, 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,
+        )
+        layer.register_parameter("w13_weight_scale", w13_weight_scale)
+        layer.register_parameter("w2_weight_scale", w2_weight_scale)
+
+        extra_weight_attrs.update(
+            {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value}
+        )
+
+        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
+        set_weight_attrs(w2_weight_scale, extra_weight_attrs)
+
+        w13_input_scale = None
+        layer.register_parameter("w13_input_scale", w13_input_scale)
+
+        w2_input_scale = None
+        layer.register_parameter("w2_input_scale", w2_input_scale)
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        layer.w13_weight = Parameter(layer.w13_weight, requires_grad=False)
+        layer.w2_weight = Parameter(layer.w2_weight, requires_grad=False)
+        layer.w13_weight_scale = Parameter(
+            layer.w13_weight_scale.data, requires_grad=False
+        )
+        layer.w2_weight_scale = Parameter(
+            layer.w2_weight_scale.data, requires_grad=False
+        )
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        router_logits: torch.Tensor,
+        top_k: int,
+        renormalize: bool,
+        use_grouped_topk: bool,
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        custom_routing_function: Optional[Callable] = None,
+        correction_bias: Optional[torch.Tensor] = None,
+        activation: str = "silu",
+        inplace: bool = True,
+        no_combine: bool = False,
+    ) -> torch.Tensor:
+        from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_experts
+        from sglang.srt.layers.moe.topk import select_experts
+
+        # Expert selection
+        topk_weights, topk_ids = select_experts(
+            hidden_states=x,
+            router_logits=router_logits,
+            use_grouped_topk=use_grouped_topk,
+            top_k=top_k,
+            renormalize=renormalize,
+            topk_group=topk_group,
+            num_expert_group=num_expert_group,
+            custom_routing_function=custom_routing_function,
+            correction_bias=correction_bias,
+        )
+
+        return fused_experts(
+            x,
+            layer.w13_weight,
+            layer.w2_weight,
+            topk_weights=topk_weights,
+            topk_ids=topk_ids,
+            inplace=inplace,
+            activation=activation,
+            use_int8_w8a8=True,
+            w1_scale=(layer.w13_weight_scale),
+            w2_scale=(layer.w2_weight_scale),
+            a1_scale=layer.w13_input_scale,
+            a2_scale=layer.w2_input_scale,
+            no_combine=no_combine,
+        )

python/sglang/srt/models/deepseek_v2.py

@@ -1202,10 +1202,9 @@ class DeepseekV2ForCausalLM(nn.Module):
                             weight, weight_scale, weight_block_size
                         )
                         self_attn.w_scale = scale
-                if (
-                    hasattr(self.quant_config, "weight_block_size")
-                    and w.dtype == torch.int8
-                ):
+                if w.dtype == torch.int8:
+                    if hasattr(self.quant_config, "weight_block_size"):
+                        # block-wise int8 need it
                         weight_block_size = self.quant_config.weight_block_size
                         if weight_block_size is not None:
                             assert hasattr(self_attn.kv_b_proj, "weight_scale_inv")
@@ -1214,6 +1213,11 @@ class DeepseekV2ForCausalLM(nn.Module):
                             w = int8_block_dequant(
                                 weight, weight_scale, weight_block_size
                             ).to(torch.bfloat16)
+                    else:
+                        # channel-wise int8 need it
+                        w = w.to(torch.bfloat16) * self_attn.kv_b_proj.weight_scale.to(
+                            torch.bfloat16
+                        )
                 w_kc, w_vc = w.unflatten(
                     0, (-1, self_attn.qk_nope_head_dim + self_attn.v_head_dim)
                 ).split([self_attn.qk_nope_head_dim, self_attn.v_head_dim], dim=1)

test/srt/run_suite.py

@@ -61,6 +61,7 @@ suites = {
         "test_w8a8_quantization.py",
         "test_fp8_kernel.py",
         "test_block_int8.py",
+        "test_int8_kernel.py",
         "test_reasoning_content.py",
     ],
     "nightly": [

test/srt/test_int8_kernel.py

+import itertools
+import unittest
+
+import torch
+
+from sglang.srt.layers.activation import SiluAndMul
+from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_moe
+from sglang.srt.layers.quantization.int8_kernel import per_token_quant_int8
+
+
+def native_w8a8_per_token_matmul(A, B, As, Bs, output_dtype=torch.float16):
+    """Matrix multiplication function that supports per-token input quantization and per-column weight quantization"""
+    A = A.to(torch.float32)
+    B = B.to(torch.float32)
+
+    assert A.shape[-1] == B.shape[-1], "Dimension mismatch"
+    assert B.ndim == 2 and B.is_contiguous(), "B must be a 2D contiguous tensor"
+
+    # Reshape input
+    M = A.numel() // A.shape[-1]
+    B = B.t()  # Transpose weight matrix
+    N, K = B.shape
+    origin_C_shape = A.shape[:-1] + (K,)
+    A = A.reshape(M, N)
+
+    # As is per-token [M, 1], Bs is per-column [1, K]
+    C = torch.matmul(A, B)  # [M, K]
+    C = As * C * Bs.view(1, -1)  # Broadcast per-column scale
+
+    return C.reshape(origin_C_shape).to(output_dtype)
+
+
+def torch_w8a8_per_column_moe(a, w1, w2, w1_s, w2_s, score, topk):
+    """This function performs fused moe with per-column int8 quantization using native torch."""
+
+    B, D = a.shape
+    # Perform per-token quantization
+    a_q, a_s = per_token_quant_int8(a)
+    # Repeat tokens to match topk
+    a_q = a_q.view(B, -1, D).repeat(1, topk, 1).reshape(-1, D)
+    # Also repeat the scale
+    a_s = a_s.view(B, -1, 1).repeat(1, topk, 1).reshape(-1, 1)  # [B*topk, 1]
+
+    out = torch.zeros(B * topk, w2.shape[1], dtype=a.dtype, device=a.device)
+
+    # Calculate routing
+    score = torch.softmax(score, dim=-1, dtype=torch.float32)
+    topk_weight, topk_ids = torch.topk(score, topk)
+    topk_weight = topk_weight.view(-1)
+    topk_ids = topk_ids.view(-1)
+    # Process each expert
+    for i in range(w1.shape[0]):
+        mask = topk_ids == i
+        if mask.sum():
+            # First MLP layer: note that a_s is now per-token
+            inter_out = native_w8a8_per_token_matmul(
+                a_q[mask], w1[i], a_s[mask], w1_s[i], output_dtype=a.dtype
+            )
+            # Activation function
+            act_out = SiluAndMul().forward_native(inter_out)
+            # Quantize activation output with per-token
+            act_out_q, act_out_s = per_token_quant_int8(act_out)
+
+            # Second MLP layer
+            out[mask] = native_w8a8_per_token_matmul(
+                act_out_q, w2[i], act_out_s, w2_s[i], output_dtype=a.dtype
+            )
+    # Apply routing weights and sum
+    return (
+        out.view(B, -1, w2.shape[1]) * topk_weight.view(B, -1, 1).to(out.dtype)
+    ).sum(dim=1)
+
+
+class TestW8A8Int8FusedMoE(unittest.TestCase):
+    DTYPES = [torch.half, torch.bfloat16]
+    M = [1, 33]
+    N = [128, 1024]
+    K = [256, 4096]
+    E = [8]
+    TOP_KS = [2, 6]
+    BLOCK_SIZE = [[64, 64], [64, 128], [128, 64], [128, 128]]
+    BLOCK_SIZE = [[128, 128]]
+    SEEDS = [0]
+
+    @classmethod
+    def setUpClass(cls):
+        if not torch.cuda.is_available():
+            raise unittest.SkipTest("CUDA is not available")
+        torch.set_default_device("cuda")
+
+    def _w8a8_int8_fused_moe(self, M, N, K, E, topk, block_size, dtype, seed):
+        torch.manual_seed(seed)
+        # Initialize int8 quantization parameters
+        factor_for_scale = 1e-2
+        int8_max = 127
+        int8_min = -128
+
+        # Input tensor
+        # M * K
+        a = torch.randn((M, K), dtype=dtype) / 10
+
+        # Generate int8 weights
+        w1_fp32 = (torch.rand((E, 2 * N, K), dtype=torch.float32) - 0.5) * 2
+        w1 = (w1_fp32 * int8_max).clamp(min=int8_min, max=int8_max).to(torch.int8)
+
+        w2_fp32 = (torch.rand((E, K, N), dtype=torch.float32) - 0.5) * 2
+        w2 = (w2_fp32 * int8_max).clamp(min=int8_min, max=int8_max).to(torch.int8)
+
+        # Generate scale for each column (per-column quantization)
+        w1_s = torch.rand(E, 2 * N, device=w1_fp32.device) * factor_for_scale
+        w2_s = torch.rand(E, K, device=w2_fp32.device) * factor_for_scale
+        score = torch.randn((M, E), dtype=dtype)
+
+        with torch.inference_mode():
+            ref_out = torch_w8a8_per_column_moe(a, w1, w2, w1_s, w2_s, score, topk)
+            out = fused_moe(
+                a,
+                w1,
+                w2,
+                score,
+                topk,
+                renormalize=False,
+                use_fp8_w8a8=False,  # Not using fp8
+                use_int8_w8a16=False,  # Not using int8-w8a16
+                use_int8_w8a8=True,  # Using int8-w8a8
+                w1_scale=w1_s,
+                w2_scale=w2_s,
+                block_shape=None,  # Not using block quantization
+            )
+
+        # Check results
+        self.assertTrue(
+            torch.mean(torch.abs(out.to(torch.float32) - ref_out.to(torch.float32)))
+            / torch.mean(torch.abs(ref_out.to(torch.float32)))
+            < 0.05
+        )
+
+    def test_w8a8_int8_fused_moe(self):
+        for params in itertools.product(
+            self.M,
+            self.N,
+            self.K,
+            self.E,
+            self.TOP_KS,
+            self.BLOCK_SIZE,
+            self.DTYPES,
+            self.SEEDS,
+        ):
+            with self.subTest(
+                M=params[0],
+                N=params[1],
+                K=params[2],
+                E=params[3],
+                topk=params[4],
+                block_size=params[5],
+                dtype=params[6],
+                seed=params[7],
+            ):
+                self._w8a8_int8_fused_moe(*params)
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)