Optimized deepseek-v3/r1 model performance on mxfp4 run (#9671)

Co-authored-by: wunhuang <wunhuang@amd.com>
Co-authored-by: wghuang <wghuang@amd.com>

python/sglang/srt/layers/communicator.py

@@ -42,10 +42,22 @@ from sglang.srt.layers.moe import (
 )
 from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
-from sglang.srt.utils import is_cuda, is_flashinfer_available, is_sm100_supported
+from sglang.srt.utils import (
+    get_bool_env_var,
+    is_cuda,
+    is_flashinfer_available,
+    is_gfx95_supported,
+    is_hip,
+    is_sm100_supported,
+)
 
 _is_flashinfer_available = is_flashinfer_available()
 _is_sm100_supported = is_cuda() and is_sm100_supported()
+_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and is_hip()
+_is_gfx95_supported = is_gfx95_supported()
+
+if _use_aiter and _is_gfx95_supported:
+    from sglang.srt.layers.quantization.rocm_mxfp4_utils import fused_rms_mxfp4_quant
 
 FUSE_ALLREDUCE_MAX_BATCH_SIZE = 2048
 
@@ -201,6 +213,7 @@ class LayerCommunicator:
         hidden_states: torch.Tensor,
         residual: torch.Tensor,
         forward_batch: ForwardBatch,
+        qaunt_format: str = "",
     ):
         if hidden_states.shape[0] == 0:
             residual = hidden_states
@@ -218,7 +231,30 @@ class LayerCommunicator:
             else:
                 if residual is None:
                     residual = hidden_states
+
+                    if _use_aiter and _is_gfx95_supported and ("mxfp4" in qaunt_format):
+                        hidden_states = fused_rms_mxfp4_quant(
+                            hidden_states,
+                            self.input_layernorm.weight,
+                            self.input_layernorm.variance_epsilon,
+                            None,
+                            None,
+                            None,
+                            None,
+                        )
+                    else:
                         hidden_states = self.input_layernorm(hidden_states)
+                else:
+                    if _use_aiter and _is_gfx95_supported and ("mxfp4" in qaunt_format):
+                        hidden_states, residual = fused_rms_mxfp4_quant(
+                            hidden_states,
+                            self.input_layernorm.weight,
+                            self.input_layernorm.variance_epsilon,
+                            None,
+                            None,
+                            None,
+                            residual,
+                        )
                     else:
                         hidden_states, residual = self.input_layernorm(
                             hidden_states, residual

python/sglang/srt/layers/quantization/quark/schemes/quark_w4a4_mxfp4.py

@@ -8,6 +8,7 @@ import torch.nn.functional as F
 from aiter.ops.gemm_op_a4w4 import gemm_a4w4
 from aiter.ops.shuffle import shuffle_weight
 from aiter.ops.triton.gemm_afp4wfp4 import gemm_afp4wfp4
+from aiter.ops.triton.gemm_afp4wfp4_pre_quant_atomic import gemm_afp4wfp4_pre_quant
 from aiter.ops.triton.quant import dynamic_mxfp4_quant
 from aiter.utility import dtypes
 from aiter.utility.fp4_utils import e8m0_shuffle
@@ -38,15 +39,6 @@ class QuarkW4A4MXFP4(QuarkScheme):
     def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
         return
 
-        # for aiter implement
-        # wshuffle = shuffle_weight(layer.weight.data, layout=(16, 16))
-        # w_scales_shuffle = e8m0_shuffle(layer.weight_scale.data).view(dtypes.fp8_e8m0)
-
-        # layer.weight = torch.nn.Parameter(wshuffle,
-        #                                  requires_grad=False)
-        # layer.weight_scale = torch.nn.Parameter(w_scales_shuffle,
-        #                                        requires_grad=False)
-
     def create_weights(
         self,
         layer: torch.nn.Module,
@@ -93,26 +85,53 @@ class QuarkW4A4MXFP4(QuarkScheme):
         x: torch.Tensor,
         bias: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
+        # This path does not have support for bias currently
+        assert bias is None, "bias is not supported"
+
+        three_d = False
+        x_s = None
+        y = None
+        if isinstance(x, tuple):
+            assert len(x) in [
+                2,
+                3,
+            ], "For tuple input, only (x, x_s) or (x, x_s, y) formats are accepted"
+            if len(x) == 2:
+                x, x_s = x
+            elif len(x) == 3:
+                x, x_s, y = x
+
+        use_fused_quant_gemm = (
+            x_s is None and y is not None and layer.weight.shape[0] == y.shape[1]
+        )
 
-        out_dtype = x.dtype
-        # M = x.shape[0]
-        # N = layer.weight.shape[0]
-
-        # quant_func = aiter.get_triton_quant(aiter.QuantType.per_1x32)
-        # x, x_scales_shuffle = quant_func(x, shuffle=True)
-
-        # y = torch.zeros((M + 255) // 256 * 256, N, device=x.device, dtype=self.out_dtype)
-
-        # out = gemm_a4w4(x, layer.weight.data, x_scales_shuffle, layer.weight_scale.data, y, bias=bias)
-
-        # return out[:M]
+        if x.dim() == 3:
+            three_d = True
+            x = x.view(-1, x.shape[-1])
+            output_shape = [*x.shape[:-1], layer.weight.shape[0]]
 
-        # triton implement
+        # use_fused_quant_gemm = true, x_q is a bf16/fp16 num
+        # x_s is not None = true, x_q is uint8 num
+        if use_fused_quant_gemm or x_s is not None:
+            x_q = x
+        else:
             x_q, x_s = dynamic_mxfp4_quant(x)
+
+        if y is None:
             y = torch.empty(
-            x_q.shape[0], layer.weight.shape[0], device=x_q.device, dtype=out_dtype
+                x_q.shape[0],
+                layer.weight.shape[0],
+                device=x_q.device,
+                dtype=self.out_dtype,
             )
 
-        out = gemm_afp4wfp4(x_q, layer.weight, x_s, layer.weight_scale, out_dtype, y)
+        if use_fused_quant_gemm:
+            gemm_afp4wfp4_pre_quant(x_q, layer.weight, layer.weight_scale, y.dtype, y)
+            y = y.to(x.dtype)
+        else:
+            gemm_afp4wfp4(x_q, layer.weight, x_s, layer.weight_scale, self.out_dtype, y)
+
+        if three_d:
+            return y.view(*output_shape)
 
-        return out
+        return y

python/sglang/srt/layers/quantization/quark/utils.py

@@ -5,6 +5,10 @@ from collections.abc import Iterable, Mapping
 from types import MappingProxyType
 from typing import Any, Optional
 
+import torch
+from aiter.ops.triton.quant import dynamic_mxfp4_quant
+from torch import nn
+
 
 def deep_compare(dict1: Any, dict2: Any) -> bool:
     if type(dict1) is not type(dict2):
@@ -105,3 +109,96 @@ def _is_equal_or_regex_match(
     elif target == value:
         return True
     return False
+
+
+# utility for tensor dims > 2 cases
+def b_dynamic_mxfp4_quant(x):
+    h, b, d = x.shape
+    x, x_scales = dynamic_mxfp4_quant(x.reshape(-1, d))
+    return x.view(h, b, d // 2), x_scales.view(h, b, d // 32)
+
+
+def mxfp4_to_f32(x, is_threed):
+    # 2 because we pack fp4 in uint8.
+    x = x.repeat_interleave(2, dim=-1)
+    if is_threed:
+        x[..., ::2] = x[..., ::2] & 0xF
+        x[..., 1::2] = x[..., 1::2] >> 4
+    else:
+        x[:, ::2] = x[:, ::2] & 0xF
+        x[:, 1::2] = x[:, 1::2] >> 4
+
+    mxfp4_list = [
+        0.0,
+        0.5,
+        1.0,
+        1.5,
+        2.0,
+        3.0,
+        4.0,
+        6.0,
+        -0.0,
+        -0.5,
+        -1.0,
+        -1.5,
+        -2.0,
+        -3.0,
+        -4.0,
+        -6.0,
+    ]
+    mxfp4_in_f32 = torch.tensor(mxfp4_list, dtype=torch.float32, device="cuda")
+    return mxfp4_in_f32[x.long()]
+
+
+def e8m0_to_f32(x):
+    # Convert the input tensor `x` (assumed to be in e8m0 format) to float32.
+    # e8m0 is a custom 8-bit floating point format with 8 bits for exponent, 0 for mantissa.
+    # This means the value is essentially 2^(exponent - 127), similar to how IEEE-754 stores floats.
+
+    # Convert x to float32 for computation, and compute the power of 2 by subtracting the bias (127).
+    x_f32 = 2 ** ((x.to(torch.float32)) - 127)
+
+    # If the exponent value was 255 (i.e., 2^(128)), this is a special case usually used to represent NaN or Inf.
+    # Since this custom format has no mantissa, treat 2^128 as NaN.
+    x_f32[x_f32 == 128] = float("nan")
+    return x_f32
+
+
+def quark_post_load_weights(self_attn: nn.Module, w: torch.Tensor, quant_format: str):
+    if "mxfp4" in quant_format:
+        # when dtype is bf16, the processing flow is to dynamic quantize bf16 tensor to uint8 tensor
+        # do w_kc (bf16) first to get the w_kc(uint8) w_s_kc(uint8)
+        # and w_vc repeating the same procedure of w_kc to get  w_vc(uint8) w_s_vc(uint8)
+        if w.dtype == 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)
+            w_kc, w_s_kc = b_dynamic_mxfp4_quant(w_kc.transpose(-2, -1))
+            w_kc = w_kc.transpose(-2, -1)
+            w_s_kc = w_s_kc.transpose(-2, -1)
+            w_vc, w_s_vc = b_dynamic_mxfp4_quant(w_vc)
+            w_s_kc = w_s_kc.transpose(1, 2).contiguous().transpose(1, 2)
+            w_s_vc = w_s_vc.contiguous().transpose(1, 2)
+        elif w.dtype == torch.uint8:  # static quant for mxfp4
+            # when dtype is uint8, it means the w has been quantized to mxfp4 format
+            # but we must separate it to w_kc and w_vc.
+            # The quantized tensor size is only half of original tensor size
+            # and the scaling factor is 1/32, the transpose behavior will be not correct
+            # need to upcast it to fp32 to separate w to w_kc and w_vc
+            # to ensure the following transpose behavior is correct
+            # and then do mxfp4 quant again
+            w = mxfp4_to_f32(w, True).to(torch.bfloat16)
+            w_scales = self_attn.kv_b_proj.weight_scale.repeat_interleave(32, dim=-1)
+            w_scales = e8m0_to_f32(w_scales).to(torch.bfloat16)
+            w = w * w_scales
+            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)
+            w_kc, w_s_kc = b_dynamic_mxfp4_quant(w_kc.transpose(-2, -1))
+            w_kc = w_kc.transpose(-2, -1)
+            w_s_kc = w_s_kc.transpose(-2, -1)
+            w_vc, w_s_vc = b_dynamic_mxfp4_quant(w_vc)
+            w_s_kc = w_s_kc.transpose(1, 2).contiguous().transpose(1, 2)
+            w_s_vc = w_s_vc.contiguous().transpose(1, 2)
+
+        return w_kc, w_s_kc, w_vc, w_s_vc

python/sglang/srt/layers/quantization/rocm_mxfp4_utils.py

+from aiter.ops.triton.batched_gemm_afp4wfp4_pre_quant import (
+    batched_gemm_afp4wfp4_pre_quant,
+)
+from aiter.ops.triton.fused_mxfp4_quant import (
+    fused_flatten_mxfp4_quant,
+    fused_rms_mxfp4_quant,
+)
+
+__all__ = [
+    "fused_rms_mxfp4_quant",
+    "fused_flatten_mxfp4_quant",
+    "batched_gemm_afp4wfp4_pre_quant",
+]

python/sglang/srt/layers/rocm_linear_utils.py

+import torch
+from aiter.ops.triton.fused_qk_concat import fused_qk_rope_cat
+from aiter.ops.triton.gemm_a16w16 import gemm_a16w16
+from aiter.ops.triton.gemm_a16w16_atomic import gemm_a16w16_atomic
+
+from sglang.srt.utils import BumpAllocator
+
+__all__ = ["fused_qk_rope_cat"]
+
+
+def aiter_dsv3_router_gemm(
+    hidden_states: torch.Tensor,
+    weight: torch.Tensor,
+    gemm_output_zero_allocator: BumpAllocator = None,
+):
+    M = hidden_states.shape[0]
+    N = weight.shape[0]
+    y = None
+
+    if M <= 256:
+        # TODO (cagri): convert to bfloat16 as part of another kernel to save time
+        # for now it is also coupled with zero allocator.
+        if gemm_output_zero_allocator != None:
+            y = gemm_output_zero_allocator.allocate(M * N).view(M, N)
+        else:
+            y = torch.zeros((M, N), dtype=torch.float32, device=hidden_states.device)
+
+    if y is not None:
+        logits = gemm_a16w16_atomic(hidden_states, weight, y=y).to(hidden_states.dtype)
+    else:
+        logits = gemm_a16w16(hidden_states, weight)
+
+    return logits
+
+
+def get_dsv3_gemm_output_zero_allocator_size(
+    n_routed_experts: int, num_moe_layers: int, allocate_size: int, embedding_dim: int
+):
+    if embedding_dim != 7168 or n_routed_experts != 256:
+        return 0
+
+    per_layer_size = 256 * (allocate_size + n_routed_experts)
+
+    return num_moe_layers * per_layer_size

python/sglang/srt/models/deepseek_v2.py

@@ -112,6 +112,7 @@ from sglang.srt.utils import (
     is_cpu,
     is_cuda,
     is_flashinfer_available,
+    is_gfx95_supported,
     is_hip,
     is_non_idle_and_non_empty,
     is_npu,
@@ -129,6 +130,22 @@ _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 _is_cpu_amx_available = cpu_has_amx_support()
 _is_cpu = is_cpu()
 _device_sm = get_device_sm()
+_is_gfx95_supported = is_gfx95_supported()
+
+_use_aiter_gfx95 = _use_aiter and _is_gfx95_supported
+
+if _use_aiter_gfx95:
+    from sglang.srt.layers.quantization.quark.utils import quark_post_load_weights
+    from sglang.srt.layers.quantization.rocm_mxfp4_utils import (
+        batched_gemm_afp4wfp4_pre_quant,
+        fused_flatten_mxfp4_quant,
+        fused_rms_mxfp4_quant,
+    )
+    from sglang.srt.layers.rocm_linear_utils import (
+        aiter_dsv3_router_gemm,
+        fused_qk_rope_cat,
+        get_dsv3_gemm_output_zero_allocator_size,
+    )
 
 if _is_cuda:
     from sgl_kernel import (
@@ -224,10 +241,17 @@ class DeepseekV2MLP(nn.Module):
         forward_batch=None,
         should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
+        gemm_output_zero_allocator: BumpAllocator = None,
     ):
         if (self.tp_size == 1) and x.shape[0] == 0:
             return x
 
+        if gemm_output_zero_allocator != None and x.shape[0] <= 256:
+            y = gemm_output_zero_allocator.allocate(
+                x.shape[0] * self.gate_up_proj.output_size_per_partition
+            ).view(x.shape[0], self.gate_up_proj.output_size_per_partition)
+            x = (x, None, y)
+
         gate_up, _ = self.gate_up_proj(x)
         x = self.act_fn(gate_up)
         x, _ = self.down_proj(
@@ -257,7 +281,7 @@ class MoEGate(nn.Module):
         if _is_cpu and _is_cpu_amx_available:
             self.quant_method = PackWeightMethod(weight_names=["weight"])
 
-    def forward(self, hidden_states):
+    def forward(self, hidden_states, gemm_output_zero_allocator: BumpAllocator = None):
         if use_intel_amx_backend(self):
             return torch.ops.sgl_kernel.weight_packed_linear(
                 hidden_states,
@@ -276,6 +300,10 @@ class MoEGate(nn.Module):
         ):
             # router gemm output float32
             logits = dsv3_router_gemm(hidden_states, self.weight)
+        elif _use_aiter_gfx95 and hidden_states.shape[0] <= 256:
+            logits = aiter_dsv3_router_gemm(
+                hidden_states, self.weight, gemm_output_zero_allocator
+            )
         else:
             logits = F.linear(hidden_states, self.weight, None)
 
@@ -439,6 +467,7 @@ class DeepseekV2MoE(nn.Module):
         forward_batch: Optional[ForwardBatch] = None,
         should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
+        gemm_output_zero_allocator: BumpAllocator = None,
     ) -> torch.Tensor:
         if not self._enable_deepep_moe:
             DUAL_STREAM_TOKEN_THRESHOLD = 1024
@@ -452,12 +481,14 @@ class DeepseekV2MoE(nn.Module):
                     hidden_states,
                     should_allreduce_fusion,
                     use_reduce_scatter,
+                    gemm_output_zero_allocator,
                 )
             else:
                 return self.forward_normal(
                     hidden_states,
                     should_allreduce_fusion,
                     use_reduce_scatter,
+                    gemm_output_zero_allocator,
                 )
         else:
             return self.forward_deepep(hidden_states, forward_batch)
@@ -467,6 +498,7 @@ class DeepseekV2MoE(nn.Module):
         hidden_states: torch.Tensor,
         should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
+        gemm_output_zero_allocator: BumpAllocator = None,
     ) -> torch.Tensor:
 
         current_stream = torch.cuda.current_stream()
@@ -475,7 +507,7 @@ class DeepseekV2MoE(nn.Module):
 
         with torch.cuda.stream(self.alt_stream):
             # router_logits: (num_tokens, n_experts)
-            router_logits = self.gate(hidden_states)
+            router_logits = self.gate(hidden_states, gemm_output_zero_allocator)
             topk_output = self.topk(hidden_states, router_logits)
             final_hidden_states = self.experts(hidden_states, topk_output)
             if not _is_cuda:
@@ -502,6 +534,7 @@ class DeepseekV2MoE(nn.Module):
         hidden_states: torch.Tensor,
         should_allreduce_fusion: bool = False,
         use_reduce_scatter: bool = False,
+        gemm_output_zero_allocator: BumpAllocator = None,
     ) -> torch.Tensor:
         if hasattr(self, "shared_experts") and use_intel_amx_backend(
             self.shared_experts.gate_up_proj
@@ -511,7 +544,7 @@ class DeepseekV2MoE(nn.Module):
         if hidden_states.shape[0] > 0:
             shared_output = self._forward_shared_experts(hidden_states)
             # router_logits: (num_tokens, n_experts)
-            router_logits = self.gate(hidden_states)
+            router_logits = self.gate(hidden_states, gemm_output_zero_allocator)
             topk_output = self.topk(hidden_states, router_logits)
         else:
             shared_output = None
@@ -1097,6 +1130,14 @@ class DeepseekV2AttentionMLA(nn.Module):
         if self.attn_mha.kv_b_proj is None:
             self.attn_mha.kv_b_proj = self.kv_b_proj
 
+        # when hidden_states is a tuple of tensors, the tuple will include quantized weight and scale tensor
+        if isinstance(hidden_states, tuple):
+            if hidden_states[0].shape[0] == 0:
+                assert (
+                    not self.o_proj.reduce_results
+                ), "short-circuiting allreduce will lead to hangs"
+                return hidden_states[0]
+        else:
             if hidden_states.shape[0] == 0:
                 assert (
                     not self.o_proj.reduce_results
@@ -1225,7 +1266,11 @@ class DeepseekV2AttentionMLA(nn.Module):
         from sglang.srt.model_executor.cuda_graph_runner import get_is_capture_mode
 
         if self.q_lora_rank is not None:
-            if hidden_states.shape[0] <= 16 and self.use_min_latency_fused_a_gemm:
+            if (
+                (not isinstance(hidden_states, tuple))
+                and hidden_states.shape[0] <= 16
+                and self.use_min_latency_fused_a_gemm
+            ):
                 fused_qkv_a_proj_out = dsv3_fused_a_gemm(
                     hidden_states, self.fused_qkv_a_proj_with_mqa.weight.T
                 )
@@ -1244,6 +1289,16 @@ class DeepseekV2AttentionMLA(nn.Module):
                 with torch.cuda.stream(self.alt_stream):
                     k_nope = self.kv_a_layernorm(k_nope)
                 current_stream.wait_stream(self.alt_stream)
+            else:
+                if _use_aiter_gfx95 and self.q_b_proj.weight.dtype == torch.uint8:
+                    q, k_nope = fused_rms_mxfp4_quant(
+                        q,
+                        self.q_a_layernorm.weight,
+                        self.q_a_layernorm.variance_epsilon,
+                        k_nope,
+                        self.kv_a_layernorm.weight,
+                        self.kv_a_layernorm.variance_epsilon,
+                    )
                 else:
                     q = self.q_a_layernorm(q)
                     k_nope = self.kv_a_layernorm(k_nope)
@@ -1278,6 +1333,23 @@ class DeepseekV2AttentionMLA(nn.Module):
             q_nope_out = q_nope_out[:, :expected_m, :]
         elif _is_hip:
             # TODO(haishaw): add bmm_fp8 to ROCm
+            if _use_aiter_gfx95 and self.w_kc.dtype == torch.uint8:
+                x = q_nope.transpose(0, 1)
+                q_nope_out = torch.empty(
+                    x.shape[0],
+                    x.shape[1],
+                    self.w_kc.shape[2],
+                    device=x.device,
+                    dtype=torch.bfloat16,
+                )
+                batched_gemm_afp4wfp4_pre_quant(
+                    x,
+                    self.w_kc.transpose(-2, -1),
+                    self.w_scale_k.transpose(-2, -1),
+                    torch.bfloat16,
+                    q_nope_out,
+                )
+            else:
                 q_nope_out = torch.bmm(
                     q_nope.to(torch.bfloat16).transpose(0, 1),
                     self.w_kc.to(torch.bfloat16) * self.w_scale,
@@ -1295,13 +1367,15 @@ class DeepseekV2AttentionMLA(nn.Module):
 
         q_nope_out = q_nope_out.transpose(0, 1)
 
-        if not self._fuse_rope_for_trtllm_mla(forward_batch):
+        if not self._fuse_rope_for_trtllm_mla(forward_batch) and (
+            not _use_aiter or not _is_gfx95_supported
+        ):
             q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
 
-        return q_pe, k_pe, q_nope_out, k_nope, forward_batch, zero_allocator
+        return q_pe, k_pe, q_nope_out, k_nope, forward_batch, zero_allocator, positions
 
     def forward_absorb_core(
-        self, q_pe, k_pe, q_nope_out, k_nope, forward_batch, zero_allocator
+        self, q_pe, k_pe, q_nope_out, k_nope, forward_batch, zero_allocator, positions
     ):
         if (
             self.current_attention_backend == "fa3"
@@ -1325,9 +1399,24 @@ class DeepseekV2AttentionMLA(nn.Module):
                 k_rope=k_pe,
                 **extra_args,
             )
+        else:
+            if _use_aiter_gfx95:
+                cos = self.rotary_emb.cos_cache
+                sin = self.rotary_emb.sin_cache
+                q, k = fused_qk_rope_cat(
+                    q_nope_out,
+                    q_pe,
+                    k_nope,
+                    k_pe,
+                    positions,
+                    cos,
+                    sin,
+                    self.rotary_emb.is_neox_style,
+                )
             else:
                 q = torch.cat([q_nope_out, q_pe], dim=-1)
                 k = torch.cat([k_nope, k_pe], dim=-1)
+
             attn_output = self.attn_mqa(q, k, k_nope, forward_batch)
         attn_output = attn_output.view(-1, self.num_local_heads, self.kv_lora_rank)
 
@@ -1352,11 +1441,34 @@ class DeepseekV2AttentionMLA(nn.Module):
             )
         elif _is_hip:
             # TODO(haishaw): add bmm_fp8 to ROCm
+            if _use_aiter_gfx95 and self.w_vc.dtype == torch.uint8:
+                x = attn_output.transpose(0, 1)
+                attn_bmm_output = torch.empty(
+                    x.shape[0],
+                    x.shape[1],
+                    self.w_vc.shape[2],
+                    device=x.device,
+                    dtype=torch.bfloat16,
+                )
+                batched_gemm_afp4wfp4_pre_quant(
+                    x,
+                    self.w_vc.transpose(-2, -1),
+                    self.w_scale_v.transpose(-2, -1),
+                    torch.bfloat16,
+                    attn_bmm_output,
+                )
+            else:
                 attn_bmm_output = torch.bmm(
                     attn_output.to(torch.bfloat16).transpose(0, 1),
                     self.w_vc.to(torch.bfloat16) * self.w_scale,
                 )
+
+            if self.o_proj.weight.dtype == torch.uint8:
+                attn_bmm_output = attn_bmm_output.transpose(0, 1)
+                attn_bmm_output = fused_flatten_mxfp4_quant(attn_bmm_output)
+            else:
                 attn_bmm_output = attn_bmm_output.transpose(0, 1).flatten(1, 2)
+
         elif self.w_vc.dtype == torch.float8_e4m3fn:
             attn_output_val, attn_output_scale = per_tensor_quant_mla_fp8(
                 attn_output.transpose(0, 1),
@@ -1864,10 +1976,21 @@ class DeepseekV2DecoderLayer(nn.Module):
         forward_batch: ForwardBatch,
         residual: Optional[torch.Tensor],
         zero_allocator: BumpAllocator,
+        gemm_output_zero_allocator: BumpAllocator = None,
     ) -> torch.Tensor:
 
+        quant_format = (
+            "mxfp4"
+            if _is_gfx95_supported
+            and self.self_attn.fused_qkv_a_proj_with_mqa.weight == torch.uint8
+            else ""
+        )
+
         hidden_states, residual = self.layer_communicator.prepare_attn(
-            hidden_states, residual, forward_batch
+            hidden_states,
+            residual,
+            forward_batch,
+            quant_format,
         )
 
         hidden_states = self.self_attn(
@@ -2036,6 +2159,37 @@ class DeepseekV2Model(nn.Module):
         else:
             self.norm = PPMissingLayer(return_tuple=True)
 
+        self.gemm_output_zero_allocator_size = 0
+        if (
+            _use_aiter_gfx95
+            and config.n_routed_experts == 256
+            and self.embed_tokens.embedding_dim == 7168
+        ):
+            num_moe_layers = sum(
+                [
+                    1
+                    for i in range(len(self.layers))
+                    if isinstance(self.layers[i].mlp, DeepseekV2MoE)
+                ]
+            )
+
+            allocate_size = 0
+            for i in range(len(self.layers)):
+                if isinstance(self.layers[i].mlp, DeepseekV2MoE):
+                    allocate_size = self.layers[
+                        i
+                    ].mlp.shared_experts.gate_up_proj.output_size_per_partition
+                    break
+
+            self.gemm_output_zero_allocator_size = (
+                get_dsv3_gemm_output_zero_allocator_size(
+                    config.n_routed_experts,
+                    num_moe_layers,
+                    allocate_size,
+                    self.embed_tokens.embedding_dim,
+                )
+            )
+
     def get_input_embeddings(self) -> torch.Tensor:
         return self.embed_tokens
 
@@ -2055,6 +2209,16 @@ class DeepseekV2Model(nn.Module):
             device=device,
         )
 
+        gemm_output_zero_allocator = (
+            BumpAllocator(
+                buffer_size=self.gemm_output_zero_allocator_size,
+                dtype=torch.float32,
+                device=device,
+            )
+            if self.gemm_output_zero_allocator_size > 0
+            else None
+        )
+
         if self.pp_group.is_first_rank:
             if input_embeds is None:
                 hidden_states = self.embed_tokens(input_ids)
@@ -2081,7 +2245,12 @@ class DeepseekV2Model(nn.Module):
             with get_global_expert_distribution_recorder().with_current_layer(i):
                 layer = self.layers[i]
                 hidden_states, residual = layer(
-                    positions, hidden_states, forward_batch, residual, zero_allocator
+                    positions,
+                    hidden_states,
+                    forward_batch,
+                    residual,
+                    zero_allocator,
+                    gemm_output_zero_allocator,
                 )
 
         if normal_end_layer != self.end_layer:
@@ -2354,6 +2523,12 @@ class DeepseekV2ForCausalLM(nn.Module):
             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)
+
+            if _use_aiter_gfx95 and self.quant_config.get_name() == "quark":
+                w_kc, self_attn.w_scale_k, w_vc, self_attn.w_scale_v = (
+                    quark_post_load_weights(self_attn, w, "mxfp4")
+                )
+
             if not use_deep_gemm_bmm:
                 self_attn.w_kc = bind_or_assign(
                     self_attn.w_kc, w_kc.transpose(1, 2).contiguous().transpose(1, 2)

python/sglang/srt/utils.py

@@ -2900,6 +2900,18 @@ def mxfp_supported():
         return False
 
 
+@lru_cache(maxsize=1)
+def is_gfx95_supported():
+    """
+    Returns whether the current platform supports MX types.
+    """
+    if torch.version.hip:
+        gcn_arch = torch.cuda.get_device_properties(0).gcnArchName
+        return any(gfx in gcn_arch for gfx in ["gfx95"])
+    else:
+        return False
+
+
 # LoRA-related constants and utilities
 SUPPORTED_LORA_TARGET_MODULES = [
     "q_proj",