Support new DeepGEMM (#7172)

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

@@ -1231,6 +1231,7 @@ class DeepEPMoE(EPMoE):
             down_input_scale,
             scale_block_size,
             masked_m,
+            scale_ue8m0=deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0,
         )
         del gateup_output
 
@@ -1238,7 +1239,13 @@ class DeepEPMoE(EPMoE):
         n = self.w2_weight.size(1)
         down_input_fp8 = (
             down_input,
-            deep_gemm_wrapper.get_col_major_tma_aligned_tensor(down_input_scale),
+            (
+                down_input_scale
+                if deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0
+                else deep_gemm_wrapper.get_col_major_tma_aligned_tensor(
+                    down_input_scale
+                )
+            ),
         )
         down_output = torch.empty(
             (num_groups, m, n), device=down_input.device, dtype=torch.bfloat16

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

@@ -584,6 +584,8 @@ class _DeepEPDispatcherImplLowLatency(_DeepEPDispatcherImplBase):
                 use_fp8=use_fp8,
                 async_finish=not self.return_recv_hook,
                 return_recv_hook=self.return_recv_hook,
+                round_scale=deep_gemm_wrapper.DEEPGEMM_V202506,
+                use_ue8m0=deep_gemm_wrapper.DEEPGEMM_V202506,
             )
         )
         return packed_recv_hidden, packed_recv_count, event, hook

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

@@ -12,6 +12,7 @@ import torch
 import triton
 import triton.language as tl
 
+from sglang.math_utils import ceil_div
 from sglang.srt.layers.moe.topk import select_experts
 from sglang.srt.layers.quantization.fp8_kernel import (
     per_token_group_quant_fp8,
@@ -518,10 +519,6 @@ def fused_moe_kernel(
     tl.store(c_ptrs, accumulator, mask=c_mask)
 
 
-def ceil_div(a, b):
-    return (a + b - 1) // b
-
-
 @triton.jit
 def moe_align_block_size_stage1(
     topk_ids_ptr,

python/sglang/srt/layers/quantization/deep_gemm_wrapper/configurer.py

@@ -21,6 +21,12 @@ def _compute_enable_deep_gemm():
 
 ENABLE_JIT_DEEPGEMM = _compute_enable_deep_gemm()
 
-DEEPGEMM_V202506 = False
+try:
+    from deep_gemm import fp8_gemm_nt
+
+    # They have not given a name to this breaking change
+    DEEPGEMM_V202506 = True
+except ImportError:
+    DEEPGEMM_V202506 = False
 
 DEEPGEMM_SCALE_UE8M0 = DEEPGEMM_V202506

python/sglang/srt/layers/quantization/deep_gemm_wrapper/entrypoint.py

@@ -16,6 +16,16 @@ logger = logging.getLogger(__name__)
 
 if ENABLE_JIT_DEEPGEMM:
     import deep_gemm
+
+    if DEEPGEMM_V202506:
+        from deep_gemm import fp8_gemm_nt as _gemm_nt_f8f8bf16_raw
+        from deep_gemm import (
+            fp8_m_grouped_gemm_nt_masked as _grouped_gemm_nt_f8f8bf16_masked_raw,
+        )
+        from deep_gemm import (
+            m_grouped_fp8_gemm_nt_contiguous as _grouped_gemm_nt_f8f8bf16_contig_raw,
+        )
+    else:
         from deep_gemm import gemm_fp8_fp8_bf16_nt as _gemm_nt_f8f8bf16_raw
         from deep_gemm import get_col_major_tma_aligned_tensor
         from deep_gemm import (

python/sglang/srt/layers/quantization/fp8_kernel.py

@@ -765,7 +765,15 @@ def prepare_block_fp8_matmul_inputs(
     assert A.shape[-1] == B.shape[-1]
     assert A.shape[:-1] == As.shape[:-1]
     assert A.is_contiguous()
+
+    if As.dtype == torch.float:
         assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1]
+    elif Bs.dtype == torch.int:
+        assert (
+            triton.cdiv(triton.cdiv(A.shape[-1], block_k), 4) == As.shape[-1]
+        ), f"{A.shape=} {As.shape=} {block_size=}"
+    else:
+        raise NotImplementedError
 
     M = A.numel() // A.shape[-1]
 
@@ -773,8 +781,17 @@ def prepare_block_fp8_matmul_inputs(
     assert B.is_contiguous()
     assert Bs.ndim == 2
     N, K = B.shape
+
+    if Bs.dtype == torch.float:
         assert triton.cdiv(N, block_n) == Bs.shape[0]
         assert triton.cdiv(K, block_k) == Bs.shape[1]
+    elif Bs.dtype == torch.int:
+        assert N == Bs.shape[0], f"{B.shape=} {Bs.shape=} {block_size=}"
+        assert (
+            triton.cdiv(triton.cdiv(K, block_k), 4) == Bs.shape[1]
+        ), f"{B.shape=} {Bs.shape=} {block_size=}"
+    else:
+        raise NotImplementedError
 
     C_shape = A.shape[:-1] + (N,)
     C = A.new_empty(C_shape, dtype=output_dtype)

python/sglang/srt/layers/quantization/fp8_utils.py

@@ -238,6 +238,7 @@ def deepgemm_w8a8_block_fp8_linear_with_fallback(
         block_size[1],
         column_major_scales=True,
         scale_tma_aligned=True,
+        scale_ue8m0=deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0,
     )
 
     if get_bool_env_var("SGLANG_W8A8_DEEPGEMM_SANITY_CHECK_UE8M0"):

python/sglang/srt/models/deepseek_v2.py

@@ -51,7 +51,7 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
+from sglang.srt.layers.moe.ep_moe.layer import DeepEPMoE, get_moe_impl_class
 from sglang.srt.layers.moe.ep_moe.token_dispatcher import DeepEPDispatcher
 from sglang.srt.layers.moe.topk import select_experts
 from sglang.srt.layers.quantization import deep_gemm_wrapper
@@ -1932,7 +1932,7 @@ class DeepseekV2ForCausalLM(nn.Module):
                 self_attn.w_vc = bind_or_assign(self_attn.w_vc, w_vc.contiguous())
                 self_attn.use_deep_gemm_bmm = True
 
-        if False:  # TODO (pr-chain)
+        if deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0:
             self._weight_requant_ue8m0()
 
     def _weight_requant_ue8m0(self):