sglang 0.4.5__py3-none-any.whl → 0.4.5.post2__py3-none-any.whl

sglang/srt/models/deepseek_v2.py

@@ -18,6 +18,8 @@
 
 import logging
 import os
+from dataclasses import dataclass
+from enum import Enum, IntEnum, auto
 from typing import Any, Dict, Iterable, Optional, Tuple
 
 import torch
@@ -27,6 +29,7 @@ from tqdm import tqdm
 from transformers import PretrainedConfig
 
 from sglang.srt.distributed import (
+    get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
     parallel_state,
     tensor_model_parallel_all_reduce,
@@ -54,9 +57,14 @@ from sglang.srt.layers.moe.ep_moe.token_dispatcher import DeepEPDispatcher
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.moe.topk import select_experts
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.quantization.fp8_kernel import (
+    _enable_jit_deepgemm_bmm,
+    per_tensor_quant_mla_deep_gemm_masked_fp8,
+    per_tensor_quant_mla_fp8,
+)
 from sglang.srt.layers.quantization.fp8_utils import (
     block_quant_to_tensor_quant,
-    input_to_float8,
+    channel_quant_to_tensor_quant,
     normalize_e4m3fn_to_e4m3fnuz,
 )
 from sglang.srt.layers.quantization.int8_utils import (
@@ -72,15 +80,16 @@ from sglang.srt.managers.expert_distribution import ExpertDistributionRecorder
 from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
 from sglang.srt.model_loader.weight_utils import default_weight_loader
-from sglang.srt.utils import DeepEPMode, add_prefix, is_cuda, is_hip
+from sglang.srt.utils import BumpAllocator, DeepEPMode, add_prefix, is_cuda, is_hip
 
 _is_hip = is_hip()
 _is_cuda = is_cuda()
 
 if _is_cuda:
-    from sgl_kernel import awq_dequantize, bmm_fp8
+    from deep_gemm import m_grouped_gemm_fp8_fp8_bf16_nt_masked
+    from sgl_kernel import awq_dequantize, bmm_fp8, merge_state_v2
 else:
-    from vllm import _custom_ops as ops
+    from vllm._custom_ops import awq_dequantize
 
 if _is_hip:
     from sglang.srt.layers.attention.triton_ops.rocm_mla_decode_rope import (
@@ -92,6 +101,18 @@ expert_distribution_recorder = ExpertDistributionRecorder()
 logger = logging.getLogger(__name__)
 
 
+class AttnForwardMethod(IntEnum):
+    # Use multi-head attention
+    MHA = auto()
+
+    # Use absorbed multi-latent attention
+    MLA = auto()
+
+    # Use multi-head attention, but with KV cache chunked.
+    # This method can avoid OOM when prefix lengths are long.
+    MHA_CHUNKED_KV = auto()
+
+
 class DeepseekV2MLP(nn.Module):
     def __init__(
         self,
@@ -131,7 +152,7 @@ class DeepseekV2MLP(nn.Module):
             )
         self.act_fn = SiluAndMul()
 
-    def forward(self, x):
+    def forward(self, x, forward_mode: Optional[ForwardMode] = None):
         gate_up, _ = self.gate_up_proj(x)
         x = self.act_fn(gate_up)
         x, _ = self.down_proj(x)
@@ -172,13 +193,8 @@ class DeepseekV2MoE(nn.Module):
         self.tp_size = get_tensor_model_parallel_world_size()
         self.routed_scaling_factor = config.routed_scaling_factor
         self.n_shared_experts = config.n_shared_experts
-        self.n_share_experts_fusion = (
-            global_server_args_dict["n_share_experts_fusion"]
-            if global_server_args_dict["n_share_experts_fusion"] is not None
-            else 0
-        )
+        self.n_share_experts_fusion = global_server_args_dict["n_share_experts_fusion"]
 
-        self.routed_scaling_factor = config.routed_scaling_factor
         if self.tp_size > config.n_routed_experts:
             raise ValueError(
                 f"Tensor parallel size {self.tp_size} is greater than "
@@ -210,6 +226,7 @@ class DeepseekV2MoE(nn.Module):
             num_expert_group=config.n_group,
             topk_group=config.topk_group,
             correction_bias=self.gate.e_score_correction_bias,
+            routed_scaling_factor=self.routed_scaling_factor,
             prefix=add_prefix("experts", prefix),
             **(
                 dict(deepep_mode=DeepEPMode[global_server_args_dict["deepep_mode"]])
@@ -278,10 +295,7 @@ class DeepseekV2MoE(nn.Module):
             return self.forward_deepep(hidden_states, forward_mode)
 
     def forward_normal(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        if self.n_shared_experts is not None and self.n_share_experts_fusion == 0:
-            shared_output = self.shared_experts(hidden_states)
-        else:
-            shared_output = None
+        shared_output = self._forward_shared_experts(hidden_states)
         # router_logits: (num_tokens, n_experts)
         router_logits = self.gate(hidden_states)
         final_hidden_states = (
@@ -311,8 +325,7 @@ class DeepseekV2MoE(nn.Module):
         ):
             # router_logits: (num_tokens, n_experts)
             router_logits = self.gate(hidden_states)
-            if self.n_shared_experts is not None:
-                shared_output = self.shared_experts(hidden_states)
+            shared_output = self._forward_shared_experts(hidden_states)
             topk_weights, topk_idx = select_experts(
                 hidden_states=hidden_states,
                 router_logits=router_logits,
@@ -322,8 +335,10 @@ class DeepseekV2MoE(nn.Module):
                 topk_group=self.topk_group,
                 num_expert_group=self.num_expert_group,
                 correction_bias=self.correction_bias,
+                routed_scaling_factor=self.routed_scaling_factor,
             )
         if self.ep_size > 1:
+            # TODO(ch-wan): allow users to set num_max_dispatch_tokens_per_rank value
             (
                 hidden_states,
                 topk_idx,
@@ -336,19 +351,15 @@ class DeepseekV2MoE(nn.Module):
                 hidden_states,
                 topk_idx,
                 topk_weights,
-                self.num_experts,
                 forward_mode=forward_mode,
             )
-        final_hidden_states = (
-            self.experts(
-                hidden_states=hidden_states,
-                reorder_topk_ids=reorder_topk_ids,
-                seg_indptr=seg_indptr,
-                masked_m=masked_m,
-                expected_m=expected_m,
-                forward_mode=forward_mode,
-            )
-            * self.routed_scaling_factor
+        final_hidden_states = self.experts(
+            hidden_states=hidden_states,
+            reorder_topk_ids=reorder_topk_ids,
+            seg_indptr=seg_indptr,
+            masked_m=masked_m,
+            expected_m=expected_m,
+            forward_mode=forward_mode,
         )
         if self.ep_size > 1:
             final_hidden_states = self.deepep_dispatcher.combine(
@@ -357,11 +368,19 @@ class DeepseekV2MoE(nn.Module):
                 topk_weights,
                 forward_mode,
             )
+        final_hidden_states *= self.routed_scaling_factor
+
         if shared_output is not None:
             final_hidden_states = final_hidden_states + shared_output
 
         return final_hidden_states
 
+    def _forward_shared_experts(self, hidden_states):
+        if self.n_share_experts_fusion == 0:
+            return self.shared_experts(hidden_states)
+        else:
+            return None
+
 
 def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
     import math
@@ -371,178 +390,6 @@ def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
     return 0.1 * mscale * math.log(scale) + 1.0
 
 
-class DeepseekV2Attention(nn.Module):
-
-    def __init__(
-        self,
-        config: PretrainedConfig,
-        hidden_size: int,
-        num_heads: int,
-        qk_nope_head_dim: int,
-        qk_rope_head_dim: int,
-        v_head_dim: int,
-        q_lora_rank: int,
-        kv_lora_rank: int,
-        rope_theta: float = 10000,
-        rope_scaling: Optional[Dict[str, Any]] = None,
-        max_position_embeddings: int = 8192,
-        quant_config: Optional[QuantizationConfig] = None,
-        layer_id=None,
-        reduce_results: bool = True,
-        prefix: str = "",
-    ) -> None:
-        super().__init__()
-        self.layer_id = layer_id
-        self.hidden_size = hidden_size
-        self.qk_nope_head_dim = qk_nope_head_dim
-        self.qk_rope_head_dim = qk_rope_head_dim
-        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
-        self.v_head_dim = v_head_dim
-        self.q_lora_rank = q_lora_rank
-        self.kv_lora_rank = kv_lora_rank
-
-        self.dp_size = get_attention_dp_size()
-        attn_tp_rank = get_attention_tp_rank()
-        attn_tp_size = get_attention_tp_size()
-
-        self.num_heads = num_heads
-        assert num_heads % attn_tp_size == 0
-        self.num_local_heads = num_heads // attn_tp_size
-        self.scaling = self.qk_head_dim**-0.5
-        self.rope_theta = rope_theta
-        self.max_position_embeddings = max_position_embeddings
-
-        if self.q_lora_rank is not None:
-            self.q_a_proj = ReplicatedLinear(
-                self.hidden_size,
-                self.q_lora_rank,
-                bias=False,
-                quant_config=quant_config,
-                prefix=add_prefix("q_a_proj", prefix),
-            )
-            self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps)
-            self.q_b_proj = ColumnParallelLinear(
-                q_lora_rank,
-                self.num_heads * self.qk_head_dim,
-                bias=False,
-                quant_config=quant_config,
-                prefix=add_prefix("q_b_proj", prefix),
-            )
-        else:
-            self.q_proj = ColumnParallelLinear(
-                self.hidden_size,
-                self.num_heads * self.qk_head_dim,
-                bias=False,
-                quant_config=quant_config,
-                prefix=add_prefix("q_proj", prefix),
-                tp_rank=attn_tp_rank,
-                tp_size=attn_tp_size,
-            )
-
-        self.kv_a_proj_with_mqa = ReplicatedLinear(
-            self.hidden_size,
-            self.kv_lora_rank + self.qk_rope_head_dim,
-            bias=False,
-            quant_config=quant_config,
-            prefix=add_prefix("kv_a_proj_with_mqa", prefix),
-        )
-        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps)
-        self.kv_b_proj = ColumnParallelLinear(
-            self.kv_lora_rank,
-            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
-            bias=False,
-            quant_config=quant_config,
-            prefix=add_prefix("kv_b_proj", prefix),
-        )
-        # O projection.
-        self.o_proj = RowParallelLinear(
-            self.num_heads * self.v_head_dim,
-            self.hidden_size,
-            bias=False,
-            quant_config=quant_config,
-            prefix=add_prefix("o_proj", prefix),
-            reduce_results=reduce_results,
-            tp_rank=attn_tp_rank,
-            tp_size=attn_tp_size,
-        )
-        rope_scaling["rope_type"] = "deepseek_yarn"
-        self.rotary_emb = get_rope_wrapper(
-            qk_rope_head_dim,
-            rotary_dim=qk_rope_head_dim,
-            max_position=max_position_embeddings,
-            base=rope_theta,
-            rope_scaling=rope_scaling,
-            is_neox_style=False,
-            device=global_server_args_dict["device"],
-        )
-
-        if rope_scaling:
-            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
-            scaling_factor = rope_scaling["factor"]
-            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
-            self.scaling = self.scaling * mscale * mscale
-
-        # TODO, support head_size 192
-        self.attn = RadixAttention(
-            self.num_local_heads,
-            256,
-            self.scaling,
-            num_kv_heads=self.num_local_heads,
-            layer_id=layer_id,
-            prefix=add_prefix("attn", prefix),
-        )
-
-    def forward(
-        self,
-        positions: torch.Tensor,
-        hidden_states: torch.Tensor,
-        forward_batch: ForwardBatch,
-    ) -> torch.Tensor:
-        if hidden_states.shape[0] == 0:
-            assert (
-                not self.o_proj.reduce_results
-            ), "short-circuiting allreduce will lead to hangs"
-            return hidden_states
-
-        if self.q_lora_rank is not None:
-            q = self.q_a_proj(hidden_states)[0]
-            q = self.q_a_layernorm(q)
-            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
-        else:
-            q = self.q_proj(hidden_states)[0].view(
-                -1, self.num_local_heads, self.qk_head_dim
-            )
-        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
-        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
-        kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
-        latent_cache = latent_cache.unsqueeze(1)
-        kv_a = self.kv_a_layernorm(kv_a.contiguous())
-        kv = self.kv_b_proj(kv_a)[0]
-        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
-        k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
-        k_pe = latent_cache[:, :, self.kv_lora_rank :]
-        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
-        q[..., self.qk_nope_head_dim :] = q_pe
-        k = torch.empty_like(q)
-        k[..., : self.qk_nope_head_dim] = k_nope
-        k[..., self.qk_nope_head_dim :] = k_pe
-        q = torch.nn.functional.pad(q, [0, 256 - self.qk_head_dim], value=0).view(
-            -1, self.num_local_heads * 256
-        )
-        k = torch.nn.functional.pad(k, [0, 256 - self.qk_head_dim], value=0).view(
-            -1, self.num_local_heads * 256
-        )
-        v = torch.nn.functional.pad(v, [0, 256 - self.v_head_dim], value=0).view(
-            -1, self.num_local_heads * 256
-        )
-        attn_output = self.attn(q, k, v, forward_batch)
-        attn_output = attn_output.view(-1, self.num_local_heads, 256)[
-            ..., : self.v_head_dim
-        ].reshape(-1, self.num_local_heads * self.v_head_dim)
-        output, _ = self.o_proj(attn_output)
-        return output
-
-
 class DeepseekV2AttentionMLA(nn.Module):
 
     def __init__(
@@ -669,6 +516,7 @@ class DeepseekV2AttentionMLA(nn.Module):
             num_kv_heads=1,
             layer_id=layer_id,
             v_head_dim=self.kv_lora_rank,
+            quant_config=quant_config,
             prefix=add_prefix("attn_mqa", prefix),
         )
 
@@ -679,6 +527,7 @@ class DeepseekV2AttentionMLA(nn.Module):
             num_kv_heads=self.num_local_heads,
             layer_id=layer_id,
             v_head_dim=self.v_head_dim,
+            quant_config=quant_config,
             prefix=add_prefix("attn_mha", prefix),
         )
 
@@ -686,39 +535,68 @@ class DeepseekV2AttentionMLA(nn.Module):
         self.w_vc = None
         self.w_scale = None
 
+        self.w_scale_k = None
+        self.w_scale_v = None
+        self.use_deep_gemm_bmm = False
+
         self.flashinfer_mla_disable_ragged = global_server_args_dict[
             "flashinfer_mla_disable_ragged"
         ]
+        self.disable_chunked_prefix_cache = global_server_args_dict[
+            "disable_chunked_prefix_cache"
+        ]
         self.attention_backend = global_server_args_dict["attention_backend"]
         self.rocm_fused_decode_mla = os.getenv("SGLANG_ROCM_FUSED_DECODE_MLA") == "1"
 
-    def no_absorb(self, forward_batch: ForwardBatch) -> bool:
+        # TODO: Design a finer way to determine the threshold
+        self.chunked_prefix_cache_threshold = 8192
+
+    def dispatch_attn_forward_method(
+        self, forward_batch: ForwardBatch
+    ) -> AttnForwardMethod:
         if self.attention_backend == "flashinfer":
             # Flashinfer MLA: Do not absorb when enabling ragged prefill
-            return (
+            if (
                 not self.flashinfer_mla_disable_ragged
                 and forward_batch.forward_mode.is_extend()
                 and not forward_batch.forward_mode.is_target_verify()
                 and not forward_batch.forward_mode.is_draft_extend()
                 and sum(forward_batch.extend_prefix_lens_cpu) == 0
-            )
+            ):
+                return AttnForwardMethod.MHA
+            else:
+                return AttnForwardMethod.MLA
         elif self.attention_backend == "fa3":
-            # Flash Attention: Keep absorbing for all extend/decode
-            return False
+            # Flash Attention: Use MHA with chunked KV cache when prefilling on long sequences.
+            if (
+                forward_batch.forward_mode.is_extend()
+                and not self.disable_chunked_prefix_cache
+                and not forward_batch.forward_mode.is_target_verify()
+                and not forward_batch.forward_mode.is_draft_extend()
+                and sum(forward_batch.extend_prefix_lens_cpu)
+                >= self.chunked_prefix_cache_threshold
+            ):
+                return AttnForwardMethod.MHA_CHUNKED_KV
+            else:
+                return AttnForwardMethod.MLA
         else:
             # Triton: Use normal computation for prefill and use weight absorption for extend/decode
-            return (
+            if (
                 forward_batch.forward_mode.is_extend()
                 and not forward_batch.forward_mode.is_target_verify()
                 and not forward_batch.forward_mode.is_draft_extend()
                 and sum(forward_batch.extend_prefix_lens_cpu) == 0
-            )
+            ):
+                return AttnForwardMethod.MHA
+            else:
+                return AttnForwardMethod.MLA
 
     def forward(
         self,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         forward_batch: ForwardBatch,
+        zero_allocator: BumpAllocator,
     ) -> torch.Tensor:
         if hidden_states.shape[0] == 0:
             assert (
@@ -726,8 +604,14 @@ class DeepseekV2AttentionMLA(nn.Module):
             ), "short-circuiting allreduce will lead to hangs"
             return hidden_states
 
-        if self.no_absorb(forward_batch):
+        attn_forward_method = self.dispatch_attn_forward_method(forward_batch)
+
+        if attn_forward_method == AttnForwardMethod.MHA:
             return self.forward_normal(positions, hidden_states, forward_batch)
+        elif attn_forward_method == AttnForwardMethod.MHA_CHUNKED_KV:
+            return self.forward_normal_chunked_kv(
+                positions, hidden_states, forward_batch
+            )
         else:
             if _is_hip:
                 if (
@@ -738,9 +622,13 @@ class DeepseekV2AttentionMLA(nn.Module):
                         positions, hidden_states, forward_batch
                     )
                 else:
-                    return self.forward_absorb(positions, hidden_states, forward_batch)
+                    return self.forward_absorb(
+                        positions, hidden_states, forward_batch, zero_allocator
+                    )
             else:
-                return self.forward_absorb(positions, hidden_states, forward_batch)
+                return self.forward_absorb(
+                    positions, hidden_states, forward_batch, zero_allocator
+                )
 
     def forward_normal(
         self,
@@ -789,6 +677,7 @@ class DeepseekV2AttentionMLA(nn.Module):
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         forward_batch: ForwardBatch,
+        zero_allocator: BumpAllocator,
     ) -> torch.Tensor:
         q_len = hidden_states.shape[0]
         q_input = hidden_states.new_empty(
@@ -804,15 +693,33 @@ class DeepseekV2AttentionMLA(nn.Module):
             )
         q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
 
-        if self.w_kc.dtype == torch.float8_e4m3fnuz:
+        if self.use_deep_gemm_bmm:
+            q_nope_val, q_nope_scale, masked_m, expected_m, aligned_m = (
+                per_tensor_quant_mla_deep_gemm_masked_fp8(
+                    q_nope.transpose(0, 1), dtype=torch.float8_e4m3fn
+                )
+            )
+            q_nope_out = q_nope.new_empty(
+                (self.num_local_heads, aligned_m, self.kv_lora_rank)
+            )
+            m_grouped_gemm_fp8_fp8_bf16_nt_masked(
+                (q_nope_val, q_nope_scale),
+                (self.w_kc, self.w_scale_k),
+                q_nope_out,
+                masked_m,
+                expected_m,
+            )
+            q_nope_out = q_nope_out[:, :expected_m, :]
+        elif self.w_kc.dtype == torch.float8_e4m3fnuz:
             # TODO(kernel): add bmm_fp8 for torch.float8_e4m3fnuz
             q_nope_out = torch.bmm(
                 q_nope.to(torch.bfloat16).transpose(0, 1),
                 self.w_kc.to(torch.bfloat16) * self.w_scale,
             )
         elif self.w_kc.dtype == torch.float8_e4m3fn:
-            q_nope_val, q_nope_scale = input_to_float8(
-                q_nope.transpose(0, 1), torch.float8_e4m3fn
+            q_nope_val, q_nope_scale = per_tensor_quant_mla_fp8(
+                q_nope.transpose(0, 1),
+                zero_allocator.allocate(1),
             )
             q_nope_out = bmm_fp8(
                 q_nope_val, self.w_kc, q_nope_scale, self.w_scale, torch.bfloat16
@@ -835,15 +742,33 @@ class DeepseekV2AttentionMLA(nn.Module):
         attn_output = self.attn_mqa(q_input, k_input, v_input, forward_batch)
         attn_output = attn_output.view(-1, self.num_local_heads, self.kv_lora_rank)
 
-        if self.w_vc.dtype == torch.float8_e4m3fnuz:
+        if self.use_deep_gemm_bmm:
+            attn_output_val, attn_output_scale, masked_m, expected_m, aligned_m = (
+                per_tensor_quant_mla_deep_gemm_masked_fp8(
+                    attn_output.transpose(0, 1), dtype=torch.float8_e4m3fn
+                )
+            )
+            attn_bmm_output = attn_output.new_empty(
+                (self.num_local_heads, aligned_m, self.v_head_dim)
+            )
+            m_grouped_gemm_fp8_fp8_bf16_nt_masked(
+                (attn_output_val, attn_output_scale),
+                (self.w_vc, self.w_scale_v),
+                attn_bmm_output,
+                masked_m,
+                expected_m,
+            )
+            attn_bmm_output = attn_bmm_output[:, :expected_m, :]
+        elif self.w_vc.dtype == torch.float8_e4m3fnuz:
             # TODO(kernel): add bmm_fp8 for torch.float8_e4m3fnuz
             attn_bmm_output = torch.bmm(
                 attn_output.to(torch.bfloat16).transpose(0, 1),
                 self.w_vc.to(torch.bfloat16) * self.w_scale,
             )
         elif self.w_vc.dtype == torch.float8_e4m3fn:
-            attn_output_val, attn_output_scale = input_to_float8(
-                attn_output.transpose(0, 1), torch.float8_e4m3fn
+            attn_output_val, attn_output_scale = per_tensor_quant_mla_fp8(
+                attn_output.transpose(0, 1),
+                zero_allocator.allocate(1),
             )
             attn_bmm_output = bmm_fp8(
                 attn_output_val,
@@ -864,6 +789,7 @@ class DeepseekV2AttentionMLA(nn.Module):
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         forward_batch: ForwardBatch,
+        zero_allocator: BumpAllocator,
     ) -> torch.Tensor:
         enable_rope_fusion = (
             os.getenv("SGLANG_FUSED_MLA_ENABLE_ROPE_FUSION", "1") == "1"
@@ -889,8 +815,10 @@ class DeepseekV2AttentionMLA(nn.Module):
                 self.w_kc.to(torch.bfloat16) * self.w_scale,
             )
         elif self.w_kc.dtype == torch.float8_e4m3fn:
-            q_nope_val, q_nope_scale = input_to_float8(
-                q_nope.transpose(0, 1), torch.float8_e4m3fn
+            q_nope_val, q_nope_scale = per_tensor_quant_mla_fp8(
+                q_nope.transpose(0, 1),
+                zero_allocator.allocate(1),
+                dtype=torch.float8_e4m3fn,
             )
             q_nope_out = bmm_fp8(
                 q_nope_val, self.w_kc, q_nope_scale, self.w_scale, torch.bfloat16
@@ -985,8 +913,10 @@ class DeepseekV2AttentionMLA(nn.Module):
                 self.w_vc.to(torch.bfloat16) * self.w_scale,
             )
         elif self.w_vc.dtype == torch.float8_e4m3fn:
-            attn_output_val, attn_output_scale = input_to_float8(
-                attn_output.transpose(0, 1), torch.float8_e4m3fn
+            attn_output_val, attn_output_scale = per_tensor_quant_mla_fp8(
+                attn_output.transpose(0, 1),
+                zero_allocator.allocate(1),
+                dtype=torch.float8_e4m3fn,
             )
             attn_bmm_output = bmm_fp8(
                 attn_output_val,
@@ -1002,6 +932,140 @@ class DeepseekV2AttentionMLA(nn.Module):
 
         return output
 
+    def _chunked_prefix_attn_mha(
+        self,
+        q: torch.Tensor,
+        accum_output: torch.Tensor,
+        accum_lse: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+
+        assert forward_batch.num_prefix_chunks is not None
+        for i in range(forward_batch.num_prefix_chunks):
+            forward_batch.set_prefix_chunk_idx(i)
+
+            # Fetch latent cache from memory pool with precomputed chunked kv indices
+            latent_cache_buf = forward_batch.token_to_kv_pool.get_key_buffer(
+                self.attn_mha.layer_id
+            )
+            latent_cache = latent_cache_buf[
+                forward_batch.prefix_chunk_kv_indices[i]
+            ].contiguous()
+
+            kv_a_normed, k_pe = latent_cache.split(
+                [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
+            )
+            kv_a_normed = kv_a_normed.squeeze(1).contiguous()
+            kv = self.kv_b_proj(kv_a_normed)[0]
+            kv = kv.view(
+                -1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim
+            )
+            v = kv[..., self.qk_nope_head_dim :]
+            k_nope = kv[..., : self.qk_nope_head_dim]
+
+            k = torch.empty(
+                (
+                    k_nope.shape[0],
+                    self.num_local_heads,
+                    self.qk_nope_head_dim + self.qk_rope_head_dim,
+                ),
+                dtype=v.dtype,
+                device=v.device,
+            )
+            k[..., : self.qk_nope_head_dim] = k_nope
+            k[..., self.qk_nope_head_dim :] = k_pe
+
+            output, lse = self.attn_mha(q, k, v, forward_batch, save_kv_cache=False)
+            lse = torch.transpose(lse, 0, 1).contiguous()
+            tmp_output = torch.empty_like(accum_output)
+            tmp_lse = torch.empty_like(accum_lse)
+            merge_state_v2(output, lse, accum_output, accum_lse, tmp_output, tmp_lse)
+            accum_output, accum_lse = tmp_output, tmp_lse
+
+        return accum_output
+
+    def forward_normal_chunked_kv(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        # In normal mha, the k and v tensors will become overly large when the prefix length is long.
+        # To avoid this, we split the kv cache into chunks and process them one after another.
+        # Since mha is compute friendly, the for loop induced here will not introduce significant overhead.
+        # The top comments in https://github.com/vllm-project/vllm/blob/main/vllm/v1/attention/backends/mla/common.py
+        # will be helpful for understanding the purpose of this function.
+
+        # First do normal mha forward to get output for extended part
+        if self.q_lora_rank is not None:
+            q = self.q_a_proj(hidden_states)[0]
+            q = self.q_a_layernorm(q)
+            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads, self.qk_head_dim)
+        else:
+            q = self.q_proj(hidden_states)[0].view(
+                -1, self.num_local_heads, self.qk_head_dim
+            )
+        _, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
+        kv_a, _ = latent_cache.split([self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+        latent_cache = latent_cache.unsqueeze(1)
+        kv_a = self.kv_a_layernorm(kv_a.contiguous())
+        kv = self.kv_b_proj(kv_a)[0]
+        kv = kv.view(-1, self.num_local_heads, self.qk_nope_head_dim + self.v_head_dim)
+        k_nope = kv[..., : self.qk_nope_head_dim]
+        v = kv[..., self.qk_nope_head_dim :]
+        k_pe = latent_cache[:, :, self.kv_lora_rank :]
+
+        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
+        q[..., self.qk_nope_head_dim :] = q_pe
+        k = torch.empty_like(q)
+        k[..., : self.qk_nope_head_dim] = k_nope
+        k[..., self.qk_nope_head_dim :] = k_pe
+
+        latent_cache[:, :, : self.kv_lora_rank] = kv_a.unsqueeze(1)
+        latent_cache[:, :, self.kv_lora_rank :] = k_pe
+
+        # Save latent cache
+        forward_batch.token_to_kv_pool.set_kv_buffer(
+            self.attn_mha, forward_batch.out_cache_loc, latent_cache, None
+        )
+
+        # Do mha for extended part without prefix
+        forward_batch.set_attn_attend_prefix_cache(False)
+        attn_output, lse = self.attn_mha(q, k, v, forward_batch, save_kv_cache=False)
+        lse = torch.transpose(lse, 0, 1).contiguous()
+
+        # Do mha attention with chunked prefix cache if there are any sequence with prefix
+        if any(forward_batch.extend_prefix_lens_cpu):
+            # Only initialize the info once
+            if forward_batch.num_prefix_chunks is None:
+                forward_batch.prepare_chunked_prefix_cache_info(q.device)
+
+            forward_batch.set_attn_attend_prefix_cache(True)
+            attn_output = self._chunked_prefix_attn_mha(
+                q=q,
+                accum_output=attn_output,
+                accum_lse=lse,
+                forward_batch=forward_batch,
+            )
+
+        attn_output = attn_output.reshape(-1, self.num_local_heads * self.v_head_dim)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class _FFNInputMode(Enum):
+    # The MLP sublayer requires 1/tp_size tokens as input
+    SCATTERED = auto()
+    # The MLP sublayer requires all tokens as input
+    FULL = auto()
+
+
+@dataclass
+class _DecoderLayerInfo:
+    is_sparse: bool
+    ffn_input_mode: _FFNInputMode
+
 
 class DeepseekV2DecoderLayer(nn.Module):
 
@@ -1013,14 +1077,6 @@ class DeepseekV2DecoderLayer(nn.Module):
         is_nextn: bool = False,
         prefix: str = "",
     ) -> None:
-
-        def is_sparse_layer(l: int):
-            return (
-                config.n_routed_experts is not None
-                and l >= config.first_k_dense_replace
-                and l % config.moe_layer_freq == 0
-            )
-
         super().__init__()
         self.hidden_size = config.hidden_size
         rope_theta = getattr(config, "rope_theta", 10000)
@@ -1031,68 +1087,54 @@ class DeepseekV2DecoderLayer(nn.Module):
         self.dp_size = get_attention_dp_size()
         self.attn_tp_size = get_attention_tp_size()
         self.attn_tp_rank = get_attention_tp_rank()
+        self.self_attn = DeepseekV2AttentionMLA(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            qk_nope_head_dim=config.qk_nope_head_dim,
+            qk_rope_head_dim=config.qk_rope_head_dim,
+            v_head_dim=config.v_head_dim,
+            q_lora_rank=(
+                config.q_lora_rank if hasattr(config, "q_lora_rank") else None
+            ),
+            kv_lora_rank=config.kv_lora_rank,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            layer_id=layer_id,
+            reduce_results=False,
+            prefix=add_prefix("self_attn", prefix),
+        )
 
-        if not global_server_args_dict["disable_mla"]:
-            self.self_attn = DeepseekV2AttentionMLA(
-                config=config,
-                hidden_size=self.hidden_size,
-                num_heads=config.num_attention_heads,
-                qk_nope_head_dim=config.qk_nope_head_dim,
-                qk_rope_head_dim=config.qk_rope_head_dim,
-                v_head_dim=config.v_head_dim,
-                q_lora_rank=(
-                    config.q_lora_rank if hasattr(config, "q_lora_rank") else None
-                ),
-                kv_lora_rank=config.kv_lora_rank,
-                rope_theta=rope_theta,
-                rope_scaling=rope_scaling,
-                max_position_embeddings=max_position_embeddings,
-                quant_config=quant_config,
-                layer_id=layer_id,
-                reduce_results=False,
-                prefix=add_prefix("self_attn", prefix),
-            )
-        else:
-            self.self_attn = DeepseekV2Attention(
-                config=config,
-                hidden_size=self.hidden_size,
-                num_heads=config.num_attention_heads,
-                qk_nope_head_dim=config.qk_nope_head_dim,
-                qk_rope_head_dim=config.qk_rope_head_dim,
-                v_head_dim=config.v_head_dim,
-                q_lora_rank=(
-                    config.q_lora_rank if hasattr(config, "q_lora_rank") else None
-                ),
-                kv_lora_rank=config.kv_lora_rank,
-                rope_theta=rope_theta,
-                rope_scaling=rope_scaling,
-                max_position_embeddings=max_position_embeddings,
-                quant_config=quant_config,
-                layer_id=layer_id,
-                reduce_results=False,
-                prefix=add_prefix("self_attn", prefix),
-            )
+        self.info = self._compute_info(config, layer_id=layer_id, is_nextn=is_nextn)
+        previous_layer_info = self._compute_info(
+            config, layer_id=layer_id - 1, is_nextn=False
+        )
 
-        if is_nextn or is_sparse_layer(layer_id):
+        if self.info.is_sparse:
             self.mlp = DeepseekV2MoE(
                 config=config,
                 quant_config=quant_config,
                 prefix=add_prefix("mlp", prefix),
             )
-            self.is_sparse = True
         else:
+            if self._enable_moe_dense_fully_dp():
+                mlp_tp_rank, mlp_tp_size = 0, 1
+            else:
+                mlp_tp_rank, mlp_tp_size = None, None
             self.mlp = DeepseekV2MLP(
                 hidden_size=config.hidden_size,
                 intermediate_size=config.intermediate_size,
                 hidden_act=config.hidden_act,
                 quant_config=quant_config,
                 prefix=add_prefix("mlp", prefix),
+                tp_rank=mlp_tp_rank,
+                tp_size=mlp_tp_size,
             )
-            self.is_sparse = False
 
         self.input_is_scattered = (
-            is_sparse_layer(layer_id - 1)
-            and global_server_args_dict["enable_deepep_moe"]
+            previous_layer_info.ffn_input_mode == _FFNInputMode.SCATTERED
         )
         self.is_last_layer = self.layer_id == config.num_hidden_layers - 1
 
@@ -1101,28 +1143,51 @@ class DeepseekV2DecoderLayer(nn.Module):
             config.hidden_size, eps=config.rms_norm_eps
         )
 
+    @staticmethod
+    def _enable_moe_dense_fully_dp():
+        return global_server_args_dict["moe_dense_tp_size"] == 1
+
+    @staticmethod
+    def _compute_info(config: PretrainedConfig, layer_id: int, is_nextn: bool):
+        is_sparse = is_nextn or (
+            config.n_routed_experts is not None
+            and layer_id >= config.first_k_dense_replace
+            and layer_id % config.moe_layer_freq == 0
+        )
+        ffn_input_mode = (
+            _FFNInputMode.SCATTERED
+            if (global_server_args_dict["enable_deepep_moe"] and is_sparse)
+            or (DeepseekV2DecoderLayer._enable_moe_dense_fully_dp() and not is_sparse)
+            else _FFNInputMode.FULL
+        )
+        return _DecoderLayerInfo(is_sparse=is_sparse, ffn_input_mode=ffn_input_mode)
+
     def forward(
         self,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         forward_batch: ForwardBatch,
         residual: Optional[torch.Tensor],
+        zero_allocator: BumpAllocator,
     ) -> torch.Tensor:
-        if global_server_args_dict["enable_deepep_moe"] and self.is_sparse:
-            return self.forward_deepep(
-                positions, hidden_states, forward_batch, residual
+        if self.info.ffn_input_mode == _FFNInputMode.SCATTERED:
+            return self.forward_ffn_with_scattered_input(
+                positions, hidden_states, forward_batch, residual, zero_allocator
             )
-        else:
-            return self.forward_normal(
-                positions, hidden_states, forward_batch, residual
+        elif self.info.ffn_input_mode == _FFNInputMode.FULL:
+            return self.forward_ffn_with_full_input(
+                positions, hidden_states, forward_batch, residual, zero_allocator
             )
+        else:
+            raise NotImplementedError
 
-    def forward_normal(
+    def forward_ffn_with_full_input(
         self,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         forward_batch: ForwardBatch,
         residual: Optional[torch.Tensor],
+        zero_allocator: BumpAllocator,
     ) -> torch.Tensor:
 
         if hidden_states.shape[0] == 0:
@@ -1143,6 +1208,7 @@ class DeepseekV2DecoderLayer(nn.Module):
                 positions=positions,
                 hidden_states=hidden_states,
                 forward_batch=forward_batch,
+                zero_allocator=zero_allocator,
             )
 
         # Gather
@@ -1184,12 +1250,13 @@ class DeepseekV2DecoderLayer(nn.Module):
 
         return hidden_states, residual
 
-    def forward_deepep(
+    def forward_ffn_with_scattered_input(
         self,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         forward_batch: ForwardBatch,
         residual: Optional[torch.Tensor],
+        zero_allocator: BumpAllocator,
     ) -> torch.Tensor:
 
         if hidden_states.shape[0] == 0:
@@ -1215,6 +1282,7 @@ class DeepseekV2DecoderLayer(nn.Module):
             positions=positions,
             hidden_states=hidden_states,
             forward_batch=forward_batch,
+            zero_allocator=zero_allocator,
         )
 
         if self.attn_tp_size != 1:
@@ -1240,7 +1308,13 @@ class DeepseekV2DecoderLayer(nn.Module):
                 hidden_states, residual = self.post_attention_layernorm(
                     hidden_states, residual
                 )
-        hidden_states = self.mlp(hidden_states, forward_batch.forward_mode)
+
+        if not (
+            self._enable_moe_dense_fully_dp()
+            and (not self.info.is_sparse)
+            and hidden_states.shape[0] == 0
+        ):
+            hidden_states = self.mlp(hidden_states, forward_batch.forward_mode)
 
         if self.is_last_layer and self.attn_tp_size != 1:
             hidden_states += residual
@@ -1296,6 +1370,14 @@ class DeepseekV2Model(nn.Module):
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
     ) -> torch.Tensor:
+        zero_allocator = BumpAllocator(
+            # TODO for two-batch-overlap, we need a larger buffer size
+            buffer_size=len(self.layers) * 2,
+            dtype=torch.float32,
+            device=(
+                input_embeds.device if input_embeds is not None else input_ids.device
+            ),
+        )
 
         if input_embeds is None:
             hidden_states = self.embed_tokens(input_ids)
@@ -1307,7 +1389,7 @@ class DeepseekV2Model(nn.Module):
             expert_distribution_recorder.set_current_layer(i)
             layer = self.layers[i]
             hidden_states, residual = layer(
-                positions, hidden_states, forward_batch, residual
+                positions, hidden_states, forward_batch, residual, zero_allocator
             )
         if not forward_batch.forward_mode.is_idle():
             if residual is None:
@@ -1330,24 +1412,33 @@ class DeepseekV2ForCausalLM(nn.Module):
         self.tp_size = get_tensor_model_parallel_world_size()
         self.quant_config = quant_config
         self.n_share_experts_fusion = global_server_args_dict["n_share_experts_fusion"]
-        # Only Deepseek V3/R1 can use shared experts fusion optimization now.
-        if (
-            global_server_args_dict.get("disable_shared_experts_fusion", False)
-            or self.config.architectures[0] != "DeepseekV3ForCausalLM"
-            or self.config.n_routed_experts != 256
-            or self.config.routed_scaling_factor != 2.5
-        ):
-            self.n_share_experts_fusion = None
-            global_server_args_dict["n_share_experts_fusion"] = None
-            logger.info(
-                "Only Deepseek V3/R1 can use shared experts fusion optimization. Shared experts fusion optimization is disabled."
-            )
-        elif self.n_share_experts_fusion is None:
-            global_server_args_dict["n_share_experts_fusion"] = self.tp_size
-            self.n_share_experts_fusion = self.tp_size
-            logger.info(
-                f"Shared experts fusion optimization is default enabled in DeepSeek V3/R1, and n_share_experts_fusion is set to {self.tp_size}. You can tune it by setting --n_share_experts_fusion or disable it by setting --disable_shared_experts_fusion."
-            )
+        if self.n_share_experts_fusion > 0:
+            # Only Deepseek V3/R1 can use shared experts fusion optimization now.
+            if (
+                self.config.architectures[0] != "DeepseekV3ForCausalLM"
+                or self.config.n_routed_experts != 256
+            ):
+                self.n_share_experts_fusion = 0
+                global_server_args_dict["n_share_experts_fusion"] = 0
+                logger.info(
+                    "Only Deepseek V3/R1 can use shared experts fusion optimization. Shared experts fusion optimization is disabled."
+                )
+            else:
+                assert (
+                    self.n_share_experts_fusion == self.tp_size
+                ), f"Shared experts fusion optimization is enabled in DeepSeek V3/R1, set it to {self.tp_size} can get best optimized performace."
+        elif self.n_share_experts_fusion == 0:
+            if (
+                torch.cuda.get_device_capability("cuda") >= (9, 0)
+                and self.config.architectures[0] == "DeepseekV3ForCausalLM"
+                and self.config.n_routed_experts == 256
+                and (not global_server_args_dict["enable_deepep_moe"])
+            ):
+                self.n_share_experts_fusion = self.tp_size
+                global_server_args_dict["n_share_experts_fusion"] = self.tp_size
+                logger.info(
+                    "Deepseek V3/R1 with fp8 can use shared experts fusion optimization when SM version >=90. Shared experts fusion optimization is enabled."
+                )
 
         self.model = DeepseekV2Model(
             config, quant_config, prefix=add_prefix("model", prefix)
@@ -1382,35 +1473,38 @@ class DeepseekV2ForCausalLM(nn.Module):
     def post_load_weights(self):
 
         # Perform post-processing after loading weights
-
-        if not global_server_args_dict["disable_mla"]:
-            for layer_id in range(self.config.num_hidden_layers):
-                self_attn = self.model.layers[layer_id].self_attn
-                if hasattr(self_attn.kv_b_proj, "qweight"):
-                    # AWQ compatible
-                    if _is_cuda:
-                        w = awq_dequantize(
-                            self_attn.kv_b_proj.qweight,
-                            self_attn.kv_b_proj.scales,
-                            self_attn.kv_b_proj.qzeros,
-                        ).T
-                    else:
-                        w = ops.awq_dequantize(
-                            self_attn.kv_b_proj.qweight,
-                            self_attn.kv_b_proj.scales,
-                            self_attn.kv_b_proj.qzeros,
-                            0,
-                            0,
-                            0,
-                        ).T
+        for layer_id in range(self.config.num_hidden_layers):
+            self_attn = self.model.layers[layer_id].self_attn
+            if hasattr(self_attn.kv_b_proj, "qweight"):
+                # AWQ compatible
+                if _is_cuda:
+                    w = awq_dequantize(
+                        self_attn.kv_b_proj.qweight,
+                        self_attn.kv_b_proj.scales,
+                        self_attn.kv_b_proj.qzeros,
+                    ).T
                 else:
-                    w = self_attn.kv_b_proj.weight
-                # NOTE(HandH1998): Since `bmm_fp8` only supports per-tensor scale, we have to requantize `self_attn.kv_b_proj`.
-                # This may affect the accuracy of fp8 model.
-                if hasattr(self.quant_config, "weight_block_size") and w.dtype in (
-                    torch.float8_e4m3fn,
-                    torch.float8_e4m3fnuz,
-                ):
+                    w = awq_dequantize(
+                        self_attn.kv_b_proj.qweight,
+                        self_attn.kv_b_proj.scales,
+                        self_attn.kv_b_proj.qzeros,
+                        0,
+                        0,
+                        0,
+                    ).T
+            else:
+                w = self_attn.kv_b_proj.weight
+            # NOTE(HandH1998): Since `bmm_fp8` only supports per-tensor scale, we have to requantize `self_attn.kv_b_proj`.
+            # This may affect the accuracy of fp8 model.
+            # Fix deepseek v3 blockwise bmm by using deep_gemm
+            use_deep_gemm_bmm = False
+            model_dtype = torch.get_default_dtype()
+
+            if w.dtype in (
+                torch.float8_e4m3fn,
+                torch.float8_e4m3fnuz,
+            ):
+                if hasattr(self.quant_config, "weight_block_size"):
                     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")
@@ -1424,29 +1518,47 @@ class DeepseekV2ForCausalLM(nn.Module):
                             weight = w
                             weight_scale = self_attn.kv_b_proj.weight_scale_inv
 
-                        w, scale = block_quant_to_tensor_quant(
-                            weight, weight_scale, weight_block_size
-                        )
-                        self_attn.w_scale = scale
-                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")
-                            weight = w
-                            weight_scale = self_attn.kv_b_proj.weight_scale_inv
-                            w = int8_block_dequant(
+                        if (
+                            _is_cuda
+                            and _enable_jit_deepgemm_bmm
+                            and weight_block_size[0] == 128
+                            and weight_block_size[1] == 128
+                            and model_dtype == torch.bfloat16
+                        ):
+                            block_scale = weight_scale
+                            use_deep_gemm_bmm = True
+                        else:
+                            w, scale = block_quant_to_tensor_quant(
                                 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)
+                            )
+                            self_attn.w_scale = scale
+                else:
+                    weight = w
+                    weight_scale = self_attn.kv_b_proj.weight_scale
+                    w, scale = channel_quant_to_tensor_quant(weight, weight_scale)
+                    self_attn.w_scale = scale
+
+            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")
+                        weight = w
+                        weight_scale = self_attn.kv_b_proj.weight_scale_inv
+                        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)
+            if not use_deep_gemm_bmm:
                 self_attn.w_kc = w_kc.transpose(1, 2).contiguous().transpose(1, 2)
                 self_attn.w_vc = w_vc.contiguous().transpose(1, 2)
                 if (
@@ -1456,6 +1568,17 @@ class DeepseekV2ForCausalLM(nn.Module):
                     self_attn.w_scale = self_attn.kv_b_proj.weight_scale
                     if _is_hip:
                         self_attn.w_scale *= 2.0
+            else:
+                num_tiles_k = self_attn.qk_nope_head_dim // weight_block_size[1]
+                num_tiles_n = self_attn.v_head_dim // weight_block_size[0]
+                ws_kc, ws_vc = block_scale.unflatten(
+                    0, (-1, (num_tiles_k + num_tiles_n))
+                ).split([num_tiles_k, num_tiles_n], dim=1)
+                self_attn.w_scale_k = ws_kc.transpose(1, 2).contiguous()
+                self_attn.w_scale_v = ws_vc.contiguous()
+                self_attn.w_kc = w_kc.transpose(1, 2).contiguous()
+                self_attn.w_vc = w_vc.contiguous()
+                self_attn.use_deep_gemm_bmm = True
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
@@ -1463,17 +1586,27 @@ class DeepseekV2ForCausalLM(nn.Module):
             ("gate_up_proj", "gate_proj", 0),
             ("gate_up_proj", "up_proj", 1),
         ]
-        if self.n_share_experts_fusion is not None and self.n_share_experts_fusion > 0:
+        if self.n_share_experts_fusion > 0:
             weights_list = list(weights)
             weights_dict = dict(weights_list)
-            suffix_list = [
-                "down_proj.weight",
-                "down_proj.weight_scale_inv",
-                "gate_proj.weight",
-                "gate_proj.weight_scale_inv",
-                "up_proj.weight",
-                "up_proj.weight_scale_inv",
-            ]
+            if self.quant_config.get_name() == "w8a8_int8":
+                suffix_list = [
+                    "down_proj.weight",
+                    "down_proj.weight_scale",
+                    "gate_proj.weight",
+                    "gate_proj.weight_scale",
+                    "up_proj.weight",
+                    "up_proj.weight_scale",
+                ]
+            else:
+                suffix_list = [
+                    "down_proj.weight",
+                    "down_proj.weight_scale_inv",
+                    "gate_proj.weight",
+                    "gate_proj.weight_scale_inv",
+                    "up_proj.weight",
+                    "up_proj.weight_scale_inv",
+                ]
             names_to_remove = []
             for moe_layer in tqdm(
                 range(
@@ -1512,12 +1645,7 @@ class DeepseekV2ForCausalLM(nn.Module):
             ckpt_gate_proj_name="gate_proj",
             ckpt_down_proj_name="down_proj",
             ckpt_up_proj_name="up_proj",
-            num_experts=self.config.n_routed_experts
-            + (
-                self.n_share_experts_fusion
-                if self.n_share_experts_fusion is not None
-                else 0
-            ),
+            num_experts=self.config.n_routed_experts + self.n_share_experts_fusion,
         )
 
         params_dict = dict(self.named_parameters())