sglang 0.4.6.post2__py3-none-any.whl → 0.4.6.post4__py3-none-any.whl

sglang/srt/layers/moe/ep_moe/layer.py

@@ -4,11 +4,19 @@ from typing import Callable, List, Optional, Tuple
 import torch
 from torch.nn import Module
 
+from sglang.srt.layers.quantization.deep_gemm import _ENABLE_JIT_DEEPGEMM
+
 try:
     from deep_gemm import (
         get_col_major_tma_aligned_tensor,
+        m_grouped_gemm_fp8_fp8_bf16_nt_contiguous,
         m_grouped_gemm_fp8_fp8_bf16_nt_masked,
     )
+    from sgl_kernel import silu_and_mul
+
+    from sglang.srt.layers.quantization.fp8_kernel import (
+        sglang_per_token_group_quant_fp8,
+    )
 
     use_deep_gemm = True
 except ImportError:
@@ -20,6 +28,8 @@ from sglang.srt.distributed import (
     get_tensor_model_parallel_world_size,
 )
 from sglang.srt.layers.moe.ep_moe.kernels import (
+    ep_gather,
+    ep_scatter,
     gelu_and_mul_triton_kernel,
     grouped_gemm_triton,
     post_reorder_triton_kernel,
@@ -27,6 +37,7 @@ from sglang.srt.layers.moe.ep_moe.kernels import (
     run_moe_ep_preproess,
     silu_and_mul_masked_post_quant_fwd,
     silu_and_mul_triton_kernel,
+    tma_align_input_scale,
 )
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
 from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoEMethodBase
@@ -600,7 +611,7 @@ class Fp8EPMoEMethod(Fp8MoEMethod):
                 self.quant_config.weight_block_size[1],
             )
             # NOTE(HandH1998): To ensure proper alignment of the block-wise quantization scales, the output_size of the weights for both the gate and up layers must be divisible by block_n.
-            # Required by collum parallel or enabling merged weights
+            # Required by column parallel or enabling merged weights
             if intermediate_size % block_n != 0:
                 raise ValueError(
                     f"The output_size of gate's and up's weight = "
@@ -842,15 +853,23 @@ class DeepEPMoE(EPMoE):
     def forward(
         self,
         hidden_states: torch.Tensor,
+        topk_idx: torch.Tensor,
+        topk_weights: torch.Tensor,
         reorder_topk_ids: torch.Tensor,
         seg_indptr: torch.Tensor,
         masked_m: torch.Tensor,
         expected_m: int,
+        num_recv_tokens_per_expert: List[int],
         forward_mode: ForwardMode,
     ):
         resolved_deepep_mode = self.deepep_mode.resolve(forward_mode)
         if resolved_deepep_mode == DeepEPMode.normal:
-            return self.forward_normal(hidden_states, reorder_topk_ids, seg_indptr)
+            if _ENABLE_JIT_DEEPGEMM:
+                return self.forward_deepgemm_contiguous(
+                    hidden_states, topk_idx, topk_weights, num_recv_tokens_per_expert
+                )
+            else:
+                return self.forward_normal(hidden_states, reorder_topk_ids, seg_indptr)
         elif resolved_deepep_mode == DeepEPMode.low_latency:
             return self.forward_deepgemm_masked(hidden_states, masked_m, expected_m)
         else:
@@ -969,6 +988,106 @@ class DeepEPMoE(EPMoE):
             )
         return down_output
 
+    def forward_deepgemm_contiguous(
+        self,
+        hidden_states_fp8: Tuple[torch.Tensor, torch.Tensor],
+        topk_idx,
+        topk_weights,
+        num_recv_tokens_per_expert: List[int],
+    ):
+        hidden_states_fp8, hidden_states_scale = hidden_states_fp8
+        assert self.quant_method is not None
+        assert self.activation == "silu"
+        if num_recv_tokens_per_expert is None:
+            return hidden_states_fp8.bfloat16()
+        all_tokens = sum(num_recv_tokens_per_expert)
+        if all_tokens <= 0:
+            return hidden_states_fp8.bfloat16()
+        M, K = hidden_states_fp8.size()
+        N = self.w13_weight.size(1)
+        scale_block_size = 128
+
+        gather_out = torch.empty_like(
+            hidden_states_fp8,
+            device=hidden_states_fp8.device,
+            dtype=torch.bfloat16,
+        )
+
+        input_tensor = [
+            torch.empty(
+                (all_tokens, K),
+                device=hidden_states_fp8.device,
+                dtype=hidden_states_fp8.dtype,
+            ),
+            torch.empty(
+                (all_tokens, K // 128),
+                device=hidden_states_fp8.device,
+                dtype=torch.float32,
+            ),
+        ]
+        m_indices = torch.empty(
+            all_tokens, device=hidden_states_fp8.device, dtype=torch.int32
+        )
+        output_index = torch.empty_like(topk_idx)
+
+        num_recv_tokens_per_expert_gpu = torch.tensor(
+            num_recv_tokens_per_expert,
+            dtype=torch.int32,
+            pin_memory=True,
+            device="cpu",
+        ).cuda(non_blocking=True)
+        expert_start_loc = torch.empty_like(num_recv_tokens_per_expert_gpu)
+
+        ep_scatter(
+            hidden_states_fp8,
+            hidden_states_scale,
+            topk_idx,
+            num_recv_tokens_per_expert_gpu,
+            expert_start_loc,
+            input_tensor[0],
+            input_tensor[1],
+            m_indices,
+            output_index,
+        )
+
+        gateup_output = torch.empty(
+            (all_tokens, N),
+            device=hidden_states_fp8.device,
+            dtype=torch.bfloat16,
+        )
+        input_tensor[1] = tma_align_input_scale(input_tensor[1])
+        m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(
+            input_tensor, self.w13_weight_fp8, gateup_output, m_indices
+        )
+        down_input = torch.empty(
+            (
+                all_tokens,
+                N // 2,
+            ),
+            device=gateup_output.device,
+            dtype=torch.bfloat16,
+        )
+        silu_and_mul(gateup_output.view(-1, N), down_input)
+        down_output = torch.empty(
+            (all_tokens, K),
+            device=hidden_states_fp8.device,
+            dtype=torch.bfloat16,
+        )
+        down_input_fp8, down_input_scale = sglang_per_token_group_quant_fp8(
+            down_input, scale_block_size
+        )
+        down_input_scale = tma_align_input_scale(down_input_scale)
+        m_grouped_gemm_fp8_fp8_bf16_nt_contiguous(
+            (down_input_fp8, down_input_scale),
+            self.w2_weight_fp8,
+            down_output,
+            m_indices,
+        )
+
+        ep_gather(down_output, topk_idx, topk_weights, output_index, gather_out)
+
+        return gather_out
+
     def forward_deepgemm_masked(
         self,
         hidden_states_fp8: Tuple[torch.Tensor, torch.Tensor],

sglang/srt/layers/moe/ep_moe/token_dispatcher.py

@@ -1,14 +1,19 @@
+from sglang.srt.layers.quantization.deep_gemm import _ENABLE_JIT_DEEPGEMM
 from sglang.srt.utils import DeepEPMode
 
 try:
     from deep_ep import Buffer
 
+    from sglang.srt.layers.quantization.fp8_kernel import (
+        sglang_per_token_group_quant_fp8,
+    )
+
     use_deepep = True
 except ImportError:
     use_deepep = False
 
 from enum import IntEnum, auto
-from typing import Optional, Tuple
+from typing import Optional, Tuple, Union
 
 import torch
 import torch.distributed as dist
@@ -78,7 +83,6 @@ class DeepEPBuffer:
                 ),
                 num_rdma_bytes,
             )
-
         cls._buffer = Buffer(
             group,
             num_nvl_bytes,
@@ -181,44 +185,74 @@ class _DeepEPDispatcherImplNormal(_DeepEPDispatcherImplBase):
         topk_weights: torch.Tensor,
     ):
         topk_idx = topk_idx.to(torch.int64)
+        if _ENABLE_JIT_DEEPGEMM:
+            # TODO hard code 128 block quant,use fp8 communication
+            hidden_states = sglang_per_token_group_quant_fp8(hidden_states, 128)
         previous_event = Buffer.capture() if self.async_finish else None
         return hidden_states, topk_idx, topk_weights, previous_event
 
     def dispatch_b(self, hidden_states, topk_idx, topk_weights, previous_event):
-        (
-            hidden_states,
-            topk_idx,
-            topk_weights,
-            event,
-        ) = self._dispatch_core(hidden_states, topk_idx, topk_weights, previous_event)
-        event.current_stream_wait() if self.async_finish else ()
-        if hidden_states.shape[0] > 0:
-            reorder_topk_ids, seg_indptr, hidden_states = self._deepep_permute(
-                hidden_states, topk_idx, fp8_dtype=hidden_states.dtype
+        if _ENABLE_JIT_DEEPGEMM:
+            (
+                hidden_states,
+                topk_idx,
+                topk_weights,
+                num_recv_tokens_per_expert_list,
+                event,
+            ) = self._dispatch_core(
+                hidden_states, topk_idx, topk_weights, previous_event
             )
-        else:
-            reorder_topk_ids = torch.empty(
-                (0,), device=hidden_states.device, dtype=torch.int64
+            event.current_stream_wait() if self.async_finish else ()
+            return (
+                hidden_states,
+                topk_idx,
+                topk_weights,
+                None,
+                num_recv_tokens_per_expert_list,
+                None,
+                None,
+                None,
             )
-            seg_indptr = torch.zeros(
-                (self.num_experts + 1,), device=hidden_states.device, dtype=torch.int64
+        else:
+            (
+                hidden_states,
+                topk_idx,
+                topk_weights,
+                num_recv_tokens_per_expert_list,
+                event,
+            ) = self._dispatch_core(
+                hidden_states, topk_idx, topk_weights, previous_event
             )
+            event.current_stream_wait() if self.async_finish else ()
+            if hidden_states.shape[0] > 0:
+                reorder_topk_ids, seg_indptr, hidden_states = self._deepep_permute(
+                    hidden_states, topk_idx, fp8_dtype=hidden_states.dtype
+                )
+            else:
+                reorder_topk_ids = torch.empty(
+                    (0,), device=hidden_states.device, dtype=torch.int64
+                )
+                seg_indptr = torch.zeros(
+                    (self.num_experts + 1,),
+                    device=hidden_states.device,
+                    dtype=torch.int64,
+                )
 
-        masked_m = expected_m = None
-
-        return (
-            hidden_states,
-            topk_idx,
-            topk_weights,
-            reorder_topk_ids,
-            seg_indptr,
-            masked_m,
-            expected_m,
-        )
+            masked_m = expected_m = None
+            return (
+                hidden_states,
+                topk_idx,
+                topk_weights,
+                reorder_topk_ids,
+                None,
+                seg_indptr,
+                masked_m,
+                expected_m,
+            )
 
     def _dispatch_core(
         self,
-        x: torch.Tensor,
+        x: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]],
         topk_idx: torch.Tensor,
         topk_weights: torch.Tensor,
         previous_event,
@@ -246,7 +280,7 @@ class _DeepEPDispatcherImplNormal(_DeepEPDispatcherImplBase):
             recv_x,
             recv_topk_idx,
             recv_topk_weights,
-            _,  # num_recv_tokens_per_expert_list
+            num_recv_tokens_per_expert_list,
             self.handle,
             event,
         ) = buffer.dispatch(
@@ -260,12 +294,14 @@ class _DeepEPDispatcherImplNormal(_DeepEPDispatcherImplBase):
             previous_event=previous_event,
             async_finish=self.async_finish,
             allocate_on_comm_stream=(previous_event is not None) and self.async_finish,
+            expert_alignment=128 if _ENABLE_JIT_DEEPGEMM else 1,
         )
 
         return (
             recv_x,
             recv_topk_idx,
             recv_topk_weights,
+            num_recv_tokens_per_expert_list,
             event,
         )
 
@@ -314,29 +350,32 @@ class _DeepEPDispatcherImplNormal(_DeepEPDispatcherImplBase):
         topk_idx: torch.Tensor,
         topk_weights: torch.Tensor,
     ):
-        if hidden_states.shape[0] > 0:
-            num_tokens = self.src2dst.shape[0] // self.router_topk
-            output = torch.empty(
-                (num_tokens, hidden_states.shape[1]),
-                device=hidden_states.device,
-                dtype=hidden_states.dtype,
-            )
-            deepep_post_reorder_triton_kernel[(num_tokens,)](
-                hidden_states,
-                output,
-                self.src2dst,
-                topk_idx,
-                topk_weights,
-                self.router_topk,
-                hidden_states.shape[1],
-                BLOCK_SIZE=512,
-            )
+        if _ENABLE_JIT_DEEPGEMM:
+            output = hidden_states
         else:
-            output = torch.zeros(
-                (0, hidden_states.shape[1]),
-                device=hidden_states.device,
-                dtype=hidden_states.dtype,
-            )
+            if hidden_states.shape[0] > 0:
+                num_tokens = self.src2dst.shape[0] // self.router_topk
+                output = torch.empty(
+                    (num_tokens, hidden_states.shape[1]),
+                    device=hidden_states.device,
+                    dtype=hidden_states.dtype,
+                )
+                deepep_post_reorder_triton_kernel[(num_tokens,)](
+                    hidden_states,
+                    output,
+                    self.src2dst,
+                    topk_idx,
+                    topk_weights,
+                    self.router_topk,
+                    hidden_states.shape[1],
+                    BLOCK_SIZE=512,
+                )
+            else:
+                output = torch.zeros(
+                    (0, hidden_states.shape[1]),
+                    device=hidden_states.device,
+                    dtype=hidden_states.dtype,
+                )
         previous_event = Buffer.capture() if self.async_finish else None
         return output, previous_event
 
@@ -360,6 +399,7 @@ class _DeepEPDispatcherImplNormal(_DeepEPDispatcherImplBase):
 
     def _get_buffer(self):
         DeepEPBuffer.set_dispatch_mode_as_normal()
+
         return DeepEPBuffer.get_deepep_buffer(
             self.group,
             self.hidden_size,
@@ -426,6 +466,7 @@ class _DeepEPDispatcherImplLowLatency(_DeepEPDispatcherImplBase):
             topk_idx,
             topk_weights,
             reorder_topk_ids,
+            None,
             seg_indptr,
             masked_m,
             expected_m,
@@ -570,7 +611,8 @@ class DeepEPDispatcher:
 
     def dispatch(self, *args, **kwargs) -> Tuple:
         self.dispatch_a(*args, **kwargs)
-        return self.dispatch_b()
+        ret = self.dispatch_b()
+        return ret
 
     def dispatch_a(
         self,
@@ -593,7 +635,8 @@ class DeepEPDispatcher:
 
     def combine(self, *args, **kwargs) -> Tuple:
         self.combine_a(*args, **kwargs)
-        return self.combine_b()
+        ret = self.combine_b()
+        return ret
 
     def combine_a(
         self,

sglang/srt/layers/moe/fused_moe_triton/configs/E=264,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    }
+}