[DeepseekV3.2] Deepseek fp8 support for MHA path (#12964)

python/sglang/srt/layers/attention/nsa_backend.py

@@ -455,7 +455,14 @@ class NativeSparseAttnBackend(AttentionBackend):
                 ]
             )
 
-            if topk_transform_method == TopkTransformMethod.RAGGED:
+            # Generate page_table_1_flattened when needed:
+            mha_dequantize_needed = (
+                self.nsa_kv_cache_store_fp8 and max_seqlen_k <= self.nsa_index_topk
+            )
+            if (
+                topk_transform_method == TopkTransformMethod.RAGGED
+                or mha_dequantize_needed
+            ):
                 page_table_1_flattened = torch.cat(
                     [
                         page_table[i, :kv_len]
@@ -468,6 +475,7 @@ class NativeSparseAttnBackend(AttentionBackend):
                     page_table_1_flattened.shape[0] == forward_batch.seq_lens_sum
                 ), f"{page_table_1_flattened.shape[0] = } must be the same as {forward_batch.seq_lens_sum = }"
 
+            if topk_transform_method == TopkTransformMethod.RAGGED:
                 topk_indices_offset = torch.repeat_interleave(
                     cu_seqlens_k[:-1],
                     forward_batch.extend_seq_lens,

python/sglang/srt/models/deepseek_v2.py

@@ -53,6 +53,7 @@ from sglang.srt.layers.attention.npu_ops.mla_preprocess import (
     NPUFusedMLAPreprocess,
     is_mla_preprocess_enabled,
 )
+from sglang.srt.layers.attention.nsa.dequant_k_cache import dequantize_k_cache_paged
 from sglang.srt.layers.attention.nsa.nsa_indexer import Indexer
 from sglang.srt.layers.attention.nsa.utils import (
     can_cp_split,
@@ -421,8 +422,6 @@ def handle_attention_nsa(attn, forward_batch):
     - Decode: MLA (avoids per-token decompression)
     - Prefill <= 2048: MHA (topk ineffective, MHA has lower FLOPs)
     - Prefill > 2048: MLA (topk filtering reduces computation significantly)
-
-    TODO: B200 (SM100) MHA path is temporarily disabled due to FA4 gpqa accuracy issues.
     """
     if forward_batch.forward_mode.is_decode_or_idle():
         return AttnForwardMethod.MLA
@@ -437,10 +436,17 @@ def handle_attention_nsa(attn, forward_batch):
         # B200 uses trtllm_ragged_attention_deepseek kernel instead of FA4
         supports_mha = _device_sm in [90, 100]
 
-        # Check if kvcache dtype is bfloat16
-        kv_dtype_is_bf16 = forward_batch.token_to_kv_pool.dtype == torch.bfloat16
+        # MHA supports both BF16 and FP8 KV cache (FP8 will be dequantized on-demand)
+        kv_dtype_supported = forward_batch.token_to_kv_pool.dtype in [
+            torch.bfloat16,
+            torch.float8_e4m3fn,
+        ]
 
-        if max_kv_len <= attn.indexer.index_topk and supports_mha and kv_dtype_is_bf16:
+        if (
+            max_kv_len <= attn.indexer.index_topk
+            and supports_mha
+            and kv_dtype_supported
+        ):
             # NSA backend uses varlen kernel which supports MHA_ONE_SHOT
             # Check if total sequence length fits in chunk capacity
             sum_seq_lens = sum(forward_batch.seq_lens_cpu)
@@ -1659,9 +1665,16 @@ class DeepseekV2AttentionMLA(nn.Module):
             forward_batch.mha_one_shot
             and sum(forward_batch.extend_prefix_lens_cpu) != 0
         ):
-            kv_a, k_pe = self._get_mla_kv_buffer(
-                forward_batch.fetch_mha_one_shot_kv_indices(), q.dtype, forward_batch
-            )
+            if self.use_nsa and self.kv_cache_dtype == "fp8_e4m3":
+                # FP8 path: dequantize NSA-specific FP8 format to BF16
+                kv_a, k_pe = self._get_mla_kv_buffer_from_fp8(forward_batch)
+            else:
+                # BF16/FP16 path: directly fetch from cache
+                kv_a, k_pe = self._get_mla_kv_buffer(
+                    forward_batch.fetch_mha_one_shot_kv_indices(),
+                    q.dtype,
+                    forward_batch,
+                )
         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]
@@ -2733,6 +2746,31 @@ class DeepseekV2AttentionMLA(nn.Module):
             kv_a = kv_a.squeeze(1).contiguous()
         return kv_a, k_pe
 
+    def _get_mla_kv_buffer_from_fp8(
+        self,
+        forward_batch: ForwardBatch,
+    ):
+        """
+        Dequantize FP8 KV cache to BF16 for MLA attention (NSA-specific format).
+
+        Returns: (kv_a, k_pe) both in BF16
+        """
+        kv_indices = forward_batch.attn_backend.forward_metadata.page_table_1_flattened
+        assert (
+            kv_indices is not None
+        ), "page_table_1_flattened should have been generated for FP8 MHA path"
+
+        kv_cache_fp8 = forward_batch.token_to_kv_pool.get_key_buffer(
+            self.attn_mha.layer_id
+        )
+
+        kv_latent_bf16 = dequantize_k_cache_paged(kv_cache_fp8, kv_indices)
+
+        kv_a = kv_latent_bf16[:, :, : self.kv_lora_rank].squeeze(1).contiguous()
+        k_pe = kv_latent_bf16[:, :, self.kv_lora_rank :]
+
+        return kv_a, k_pe
+
     def _concat_and_cast_mha_k(self, k_nope, k_pe, forward_batch):
         # Temporary for DeepSeek V3/R1 only, but can generalize if needed
         k_shape = (k_nope.shape[0], self.num_local_heads, self.qk_head_dim)