sglang 0.4.4.post3__py3-none-any.whl → 0.4.4.post4__py3-none-any.whl

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

@@ -13,11 +13,6 @@ import triton
 import triton.language as tl
 
 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,
-    per_token_quant_int8,
-)
 from sglang.srt.utils import (
     direct_register_custom_op,
     get_bool_env_var,
@@ -42,9 +37,6 @@ if _is_cuda:
     from sgl_kernel import gelu_and_mul, silu_and_mul
 
     from sglang.srt.custom_op import scaled_fp8_quant as sgl_scaled_fp8_quant
-    from sglang.srt.layers.quantization.fp8_kernel import (
-        sglang_per_token_group_quant_fp8,
-    )
 else:
     from vllm import _custom_ops as vllm_ops
 
@@ -52,6 +44,257 @@ if _is_cuda or _is_hip:
     from sgl_kernel import moe_align_block_size as sgl_moe_align_block_size
 
 
+@triton.jit
+def write_zeros_to_output(
+    c_ptr,
+    stride_cm,
+    stride_cn,
+    pid_n,
+    N,
+    offs_token,
+    token_mask,
+    BLOCK_SIZE_M,
+    BLOCK_SIZE_N,
+    compute_type,
+):
+    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=compute_type)
+    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :]
+    c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
+    tl.store(c_ptrs, accumulator, mask=c_mask)
+
+
+@triton.jit
+def fused_moe_kernel_gptq_awq(
+    # Pointers to matrices
+    a_ptr,
+    b_ptr,
+    c_ptr,
+    b_scale_ptr,
+    b_zp_ptr,
+    topk_weights_ptr,
+    sorted_token_ids_ptr,
+    expert_ids_ptr,
+    num_tokens_post_padded_ptr,
+    # Matrix dimensions
+    N: tl.constexpr,
+    K: tl.constexpr,
+    EM,
+    num_valid_tokens,
+    # The stride variables represent how much to increase the ptr by when
+    # moving by 1 element in a particular dimension. E.g. `stride_am` is
+    # how much to increase `a_ptr` by to get the element one row down
+    # (A has M rows).
+    stride_am,
+    stride_ak,
+    stride_be,
+    stride_bk,
+    stride_bn,
+    stride_cm,
+    stride_cn,
+    stride_bse,
+    stride_bsk,
+    stride_bsn,
+    stride_bze,
+    stride_bzk,
+    stride_bzn,
+    group_size: tl.constexpr,
+    # Meta-parameters
+    BLOCK_SIZE_M: tl.constexpr,
+    BLOCK_SIZE_N: tl.constexpr,
+    BLOCK_SIZE_K: tl.constexpr,
+    GROUP_SIZE_M: tl.constexpr,
+    MUL_ROUTED_WEIGHT: tl.constexpr,
+    top_k: tl.constexpr,
+    compute_type: tl.constexpr,
+    has_zp: tl.constexpr,
+    use_int4_w4a16: tl.constexpr,
+    use_int8_w8a16: tl.constexpr,
+    even_Ks: tl.constexpr,
+):
+    """
+    Implements the fused computation for a Mixture of Experts (MOE) using
+    token and expert matrices.
+    Key Parameters:
+    - A: The input tensor representing tokens with shape (*, K), where '*' can
+        be any shape representing batches and K is the feature dimension of
+        each token.
+    - B: The stacked MOE weight tensor with shape (E, N, K), where E is
+        the number of experts, K is the input feature dimension, and N is
+        the output feature dimension.
+    - C: The output cache tensor with shape (M, topk, N), where M is the
+        total number of tokens post padding, topk is the number of times
+        each token is repeated, and N is the output feature dimension.
+    - sorted_token_ids: A tensor containing the sorted indices of tokens,
+        repeated topk times and arranged by the expert index they are
+        assigned to.
+    - 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
+    BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix
+    multiplication across different blocks processed by the same expert.
+    """
+    # -----------------------------------------------------------
+    # Map program ids `pid` to the block of C it should compute.
+    # This is done in a grouped ordering to promote L2 data reuse.
+    pid = tl.program_id(axis=0)
+    num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M)
+    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
+    num_pid_in_group = GROUP_SIZE_M * num_pid_n
+    group_id = pid // num_pid_in_group
+    first_pid_m = group_id * GROUP_SIZE_M
+    group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
+    pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
+    pid_n = (pid % num_pid_in_group) // group_size_m
+
+    # ----------------------------------------------------------
+    # Create pointers for the first blocks of A and B.
+    # We will advance this pointer as we move in the K direction
+    # and accumulate
+    # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
+    # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
+    num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
+    if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
+        return
+    offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
+    offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
+    token_mask = offs_token < num_valid_tokens
+
+    off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
+    if off_experts == -1:
+        # -----------------------------------------------------------
+        # Write back zeros to the output when the expert is not
+        # in the current expert parallel rank.
+        write_zeros_to_output(
+            c_ptr,
+            stride_cm,
+            stride_cn,
+            pid_n,
+            N,
+            offs_token,
+            token_mask,
+            BLOCK_SIZE_M,
+            BLOCK_SIZE_N,
+            compute_type,
+        )
+        return
+
+    offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N
+    offs_k = tl.arange(0, BLOCK_SIZE_K)
+    a_ptrs = a_ptr + (
+        offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak
+    )
+
+    if use_int4_w4a16:
+        b_ptrs = (
+            b_ptr
+            + off_experts * stride_be
+            + (offs_k[:, None] // 2) * stride_bk
+            + offs_bn[None, :] * stride_bn
+        )
+        b_shifter = (offs_k[:, None] % 2) * 4
+    elif use_int8_w8a16:
+        b_ptrs = (
+            b_ptr
+            + off_experts * stride_be
+            + offs_k[:, None] * stride_bk
+            + offs_bn[None, :] * stride_bn
+        )
+
+    if not has_zp and use_int4_w4a16:
+        b_zp_num = 8
+    if not has_zp and use_int8_w8a16:
+        b_zp_num = 128
+    elif has_zp and use_int4_w4a16:
+        b_zp_shifter = (offs_bn[None, :] % 2) * 4
+
+    # -----------------------------------------------------------
+    # Iterate to compute a block of the C matrix.
+    # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
+    # of fp32 values for higher accuracy.
+    # `accumulator` will be converted back to fp16 after the loop.
+    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
+    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
+        # Load the next block of A and B, generate a mask by checking the
+        # K dimension.
+
+        if not even_Ks:
+            k_mask = offs_k[:, None] < K - k * BLOCK_SIZE_K
+            k_other = 0.0
+        else:
+            k_mask = None
+            k_other = None
+
+        a = tl.load(
+            a_ptrs,
+            mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K),
+            other=0.0,
+        )
+        b = tl.load(b_ptrs)
+        if use_int4_w4a16:
+            b = (b >> b_shifter) & 0xF
+
+        b_scale_ptrs = (
+            b_scale_ptr
+            + off_experts * stride_bse
+            + offs_bn[None, :] * stride_bsn
+            + ((offs_k[:, None] + BLOCK_SIZE_K * k) // group_size) * stride_bsk
+        )
+        b_scale = tl.load(b_scale_ptrs, mask=k_mask, other=k_other)
+        b_scale = b_scale.to(tl.float32)
+
+        if has_zp and use_int4_w4a16:
+            offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
+            b_zp_ptrs = (
+                b_zp_ptr
+                + off_experts * stride_bze
+                + (offs_bn[None, :] // 2) * stride_bzn
+                + offs_k_true * stride_bzk
+            )
+            b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
+            b_zp = (b_zp >> b_zp_shifter) & 0xF
+            b_zp = b_zp.to(tl.float32)
+        elif has_zp and use_int8_w8a16:
+            offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size
+            b_zp_ptrs = (
+                b_zp_ptr
+                + off_experts * stride_bze
+                + offs_bn[None, :] * stride_bzn
+                + offs_k_true * stride_bzk
+            )
+            b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other)
+            b_zp = b_zp.to(tl.float32)
+
+        # We accumulate along the K dimension.
+        if has_zp:
+            b = ((b.to(tl.float32) - b_zp) * b_scale).to(compute_type)
+        else:
+            b = ((b.to(tl.float32) - b_zp_num) * b_scale).to(compute_type)
+        accumulator = tl.dot(a, b, acc=accumulator)
+
+        # Advance the ptrs to the next K block.
+        a_ptrs += BLOCK_SIZE_K * stride_ak
+        if use_int4_w4a16:
+            b_ptrs += (BLOCK_SIZE_K // 2) * stride_bk
+        else:
+            b_ptrs += BLOCK_SIZE_K * stride_bk
+
+    if MUL_ROUTED_WEIGHT:
+        moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
+        accumulator = accumulator * moe_weight[:, None]
+
+    accumulator = accumulator.to(compute_type)
+    # -----------------------------------------------------------
+    # Write back the block of the output
+    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :]
+    c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
+    tl.store(c_ptrs, accumulator, mask=c_mask)
+
+
 @triton.jit
 def fused_moe_kernel(
     # Pointers to matrices
@@ -152,6 +395,7 @@ def fused_moe_kernel(
         return
     offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
     offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
+    offs_token = offs_token.to(tl.int64)
     token_mask = offs_token < num_valid_tokens
 
     offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
@@ -495,6 +739,7 @@ def invoke_fused_moe_kernel(
     C: torch.Tensor,
     A_scale: Optional[torch.Tensor],
     B_scale: Optional[torch.Tensor],
+    B_zp: Optional[torch.Tensor],
     topk_weights: torch.Tensor,
     topk_ids: torch.Tensor,
     sorted_token_ids: torch.Tensor,
@@ -507,9 +752,20 @@ def invoke_fused_moe_kernel(
     use_fp8_w8a8: bool,
     use_int8_w8a8: bool,
     use_int8_w8a16: bool,
+    use_int4_w4a16: bool,
     block_shape: Optional[List[int]] = None,
     no_combine: bool = False,
 ) -> None:
+    from sglang.srt.layers.quantization.int8_kernel import (
+        per_token_group_quant_int8,
+        per_token_quant_int8,
+    )
+
+    if _is_cuda:
+        from sglang.srt.layers.quantization.fp8_kernel import (
+            sglang_per_token_group_quant_fp8,
+        )
+
     assert topk_weights.stride(1) == 1
     assert sorted_token_ids.stride(0) == 1
 
@@ -547,8 +803,9 @@ def invoke_fused_moe_kernel(
             assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
             assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
             assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]
-    elif use_int8_w8a16:
+    elif use_int8_w8a16 or use_int4_w4a16:
         assert B_scale is not None
+        assert block_shape is None or block_shape[0] == 0
     else:
         assert A_scale is None
         assert B_scale is None
@@ -564,43 +821,90 @@ def invoke_fused_moe_kernel(
     else:
         even_Ks = False
 
-    fused_moe_kernel[grid](
-        A,
-        B,
-        C,
-        A_scale,
-        B_scale,
-        topk_weights,
-        sorted_token_ids,
-        expert_ids,
-        num_tokens_post_padded,
-        B.shape[1],
-        B.shape[2] - padded_size,
-        sorted_token_ids.shape[0],
-        topk_ids.numel(),
-        A.stride(0),
-        A.stride(1),
-        B.stride(0),
-        B.stride(2),
-        B.stride(1),
-        C.stride(1),
-        C.stride(2),
-        A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0,
-        A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0,
-        B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0,
-        B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0,
-        B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0,
-        0 if block_shape is None else block_shape[0],
-        0 if block_shape is None else block_shape[1],
-        MUL_ROUTED_WEIGHT=mul_routed_weight,
-        top_k=top_k,
-        compute_type=compute_type,
-        use_fp8_w8a8=use_fp8_w8a8,
-        use_int8_w8a8=use_int8_w8a8,
-        use_int8_w8a16=use_int8_w8a16,
-        even_Ks=even_Ks,
-        **config,
-    )
+    if (
+        (use_int8_w8a16 or use_int4_w4a16)
+        and block_shape is not None
+        and block_shape[1] > 0
+    ):
+        assert B_scale is not None and B_scale.ndim == 3
+        assert B_zp is None or B_zp.ndim == 3
+        fused_moe_kernel_gptq_awq[grid](
+            A,
+            B,
+            C,
+            B_scale,
+            B_zp,
+            topk_weights,
+            sorted_token_ids,
+            expert_ids,
+            num_tokens_post_padded,
+            B.shape[1],
+            A.shape[1],
+            sorted_token_ids.shape[0],
+            topk_ids.numel(),
+            A.stride(0),
+            A.stride(1),
+            B.stride(0),
+            B.stride(2),
+            B.stride(1),
+            C.stride(1),
+            C.stride(2),
+            B_scale.stride(0),
+            B_scale.stride(2),
+            B_scale.stride(1),
+            B_zp.stride(0) if B_zp is not None else 0,
+            B_zp.stride(2) if B_zp is not None else 0,
+            B_zp.stride(1) if B_zp is not None else 0,
+            group_size=block_shape[1],
+            MUL_ROUTED_WEIGHT=mul_routed_weight,
+            top_k=top_k,
+            compute_type=compute_type,
+            has_zp=B_zp is not None,
+            use_int4_w4a16=use_int4_w4a16,
+            use_int8_w8a16=use_int8_w8a16,
+            even_Ks=even_Ks,
+            **config,
+        )
+
+    else:
+
+        fused_moe_kernel[grid](
+            A,
+            B,
+            C,
+            A_scale,
+            B_scale,
+            topk_weights,
+            sorted_token_ids,
+            expert_ids,
+            num_tokens_post_padded,
+            B.shape[1],
+            B.shape[2] - padded_size,
+            sorted_token_ids.shape[0],
+            topk_ids.numel(),
+            A.stride(0),
+            A.stride(1),
+            B.stride(0),
+            B.stride(2),
+            B.stride(1),
+            C.stride(1),
+            C.stride(2),
+            A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0,
+            A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0,
+            B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0,
+            B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0,
+            B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0,
+            0 if block_shape is None else block_shape[0],
+            0 if block_shape is None else block_shape[1],
+            MUL_ROUTED_WEIGHT=mul_routed_weight,
+            top_k=top_k,
+            compute_type=compute_type,
+            use_fp8_w8a8=use_fp8_w8a8,
+            use_int8_w8a8=use_int8_w8a8,
+            use_int8_w8a16=use_int8_w8a16,
+            even_Ks=even_Ks,
+            **config,
+        )
 
 
 def get_config_file_name(
@@ -749,6 +1053,7 @@ def try_get_optimal_moe_config(
 def get_config_dtype_str(
     dtype: torch.dtype,
     use_int8_w8a16: Optional[bool] = False,
+    use_int4_w4a16: Optional[bool] = False,
     use_fp8_w8a8: Optional[bool] = False,
     use_int8_w8a8: Optional[bool] = False,
 ):
@@ -756,6 +1061,8 @@ def get_config_dtype_str(
         return "fp8_w8a8"
     elif use_int8_w8a8:
         return "int8_w8a8"
+    elif use_int4_w4a16:
+        return "int4_w4a16"
     elif use_int8_w8a16:
         return "int8_w8a16"
     elif dtype == torch.float:
@@ -775,8 +1082,11 @@ def inplace_fused_experts(
     use_fp8_w8a8: bool = False,
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
+    use_int4_w4a16: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
+    w1_zp: Optional[torch.Tensor] = None,
+    w2_zp: Optional[torch.Tensor] = None,
     a1_scale: Optional[torch.Tensor] = None,
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
@@ -792,8 +1102,11 @@ def inplace_fused_experts(
         use_fp8_w8a8,
         use_int8_w8a8,
         use_int8_w8a16,
+        use_int4_w4a16,
         w1_scale,
         w2_scale,
+        w1_zp,
+        w2_zp,
         a1_scale,
         a2_scale,
         block_shape,
@@ -810,8 +1123,11 @@ def inplace_fused_experts_fake(
     use_fp8_w8a8: bool = False,
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
+    use_int4_w4a16: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
+    w1_zp: Optional[torch.Tensor] = None,
+    w2_zp: Optional[torch.Tensor] = None,
     a1_scale: Optional[torch.Tensor] = None,
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
@@ -837,8 +1153,11 @@ def outplace_fused_experts(
     use_fp8_w8a8: bool = False,
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
+    use_int4_w4a16: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
+    w1_zp: Optional[torch.Tensor] = None,
+    w2_zp: Optional[torch.Tensor] = None,
     a1_scale: Optional[torch.Tensor] = None,
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
@@ -855,8 +1174,11 @@ def outplace_fused_experts(
         use_fp8_w8a8,
         use_int8_w8a8,
         use_int8_w8a16,
+        use_int4_w4a16,
         w1_scale,
         w2_scale,
+        w1_zp,
+        w2_zp,
         a1_scale,
         a2_scale,
         block_shape,
@@ -874,8 +1196,11 @@ def outplace_fused_experts_fake(
     use_fp8_w8a8: bool = False,
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
+    use_int4_w4a16: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
+    w1_zp: Optional[torch.Tensor] = None,
+    w2_zp: Optional[torch.Tensor] = None,
     a1_scale: Optional[torch.Tensor] = None,
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
@@ -903,8 +1228,11 @@ def fused_experts(
     use_fp8_w8a8: bool = False,
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
+    use_int4_w4a16: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
+    w1_zp: Optional[torch.Tensor] = None,
+    w2_zp: Optional[torch.Tensor] = None,
     a1_scale: Optional[torch.Tensor] = None,
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
@@ -922,8 +1250,11 @@ def fused_experts(
             use_fp8_w8a8,
             use_int8_w8a8,
             use_int8_w8a16,
+            use_int4_w4a16,
             w1_scale,
             w2_scale,
+            w1_zp,
+            w2_zp,
             a1_scale,
             a2_scale,
             block_shape,
@@ -940,8 +1271,11 @@ def fused_experts(
             use_fp8_w8a8,
             use_int8_w8a8,
             use_int8_w8a16,
+            use_int4_w4a16,
             w1_scale,
             w2_scale,
+            w1_zp,
+            w2_zp,
             a1_scale,
             a2_scale,
             block_shape,
@@ -960,8 +1294,11 @@ def fused_experts_impl(
     use_fp8_w8a8: bool = False,
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
+    use_int4_w4a16: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
+    w1_zp: Optional[torch.Tensor] = None,
+    w2_zp: Optional[torch.Tensor] = None,
     a1_scale: Optional[torch.Tensor] = None,
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
@@ -976,7 +1313,12 @@ def fused_experts_impl(
         padded_size = 0
 
     # Check constraints.
-    assert hidden_states.shape[1] == w1.shape[2] - padded_size, "Hidden size mismatch"
+    if use_int4_w4a16:
+        assert hidden_states.shape[1] // 2 == w1.shape[2], "Hidden size mismatch"
+    else:
+        assert (
+            hidden_states.shape[1] == w1.shape[2] - padded_size
+        ), "Hidden size mismatch"
     assert topk_weights.shape == topk_ids.shape, "topk shape mismatch"
     assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
     assert w1.is_contiguous(), "Expert weights1 must be contiguous"
@@ -993,6 +1335,7 @@ def fused_experts_impl(
         use_fp8_w8a8=use_fp8_w8a8,
         use_int8_w8a8=use_int8_w8a8,
         use_int8_w8a16=use_int8_w8a16,
+        use_int4_w4a16=use_int4_w4a16,
         dtype=hidden_states.dtype,
     )
 
@@ -1074,6 +1417,7 @@ def fused_experts_impl(
             intermediate_cache1,
             a1_scale,
             w1_scale,
+            w1_zp,
             curr_topk_weights,
             curr_topk_ids,
             sorted_token_ids,
@@ -1086,6 +1430,7 @@ def fused_experts_impl(
             use_fp8_w8a8=use_fp8_w8a8,
             use_int8_w8a8=use_int8_w8a8,
             use_int8_w8a16=use_int8_w8a16,
+            use_int4_w4a16=use_int4_w4a16,
             block_shape=block_shape,
         )
         if activation == "silu":
@@ -1115,6 +1460,7 @@ def fused_experts_impl(
             ),
             a2_scale,
             w2_scale,
+            w2_zp,
             curr_topk_weights,
             curr_topk_ids,
             sorted_token_ids,
@@ -1127,6 +1473,7 @@ def fused_experts_impl(
             use_fp8_w8a8=use_fp8_w8a8,
             use_int8_w8a8=use_int8_w8a8,
             use_int8_w8a16=use_int8_w8a16,
+            use_int4_w4a16=use_int4_w4a16,
             block_shape=block_shape,
         )
 
@@ -1172,8 +1519,11 @@ def fused_moe(
     use_fp8_w8a8: bool = False,
     use_int8_w8a8: bool = False,
     use_int8_w8a16: bool = False,
+    use_int4_w4a16: bool = False,
     w1_scale: Optional[torch.Tensor] = None,
     w2_scale: Optional[torch.Tensor] = None,
+    w1_zp: Optional[torch.Tensor] = None,
+    w2_zp: Optional[torch.Tensor] = None,
     a1_scale: Optional[torch.Tensor] = None,
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
@@ -1203,6 +1553,9 @@ def fused_moe(
         products for w1 and w2. Defaults to False.
     - use_int8_w8a16 (bool): If True, use fp8 arithmetic to compute the inner
         products for w1 and w2. Defaults to False.
+    - use_int4_w4a16 (bool): If True, use matmul of int4 weight and bf16/fp16
+        activation to compute the inner products for w1 and w2.
+        Defaults to False.
     - w1_scale (Optional[torch.Tensor]): Optional scale to be used for
         w1.
     - w2_scale (Optional[torch.Tensor]): Optional scale to be used for
@@ -1242,8 +1595,11 @@ def fused_moe(
         use_fp8_w8a8=use_fp8_w8a8,
         use_int8_w8a8=use_int8_w8a8,
         use_int8_w8a16=use_int8_w8a16,
+        use_int4_w4a16=use_int4_w4a16,
         w1_scale=w1_scale,
         w2_scale=w2_scale,
+        w1_zp=w1_zp,
+        w2_zp=w2_zp,
         a1_scale=a1_scale,
         a2_scale=a2_scale,
         block_shape=block_shape,