[diffusion] feat: add SageAttention packed varlen path for minimax-h3 (#33703)

python/sglang/multimodal_gen/configs/models/dits/minimax_h3.py

@@ -18,6 +18,7 @@ class MiniMaxH3DiTArchConfig(DiTArchConfig):
     _supported_attention_backends: set[AttentionBackendEnum] = field(
         default_factory=lambda: {
             AttentionBackendEnum.FA,
+            AttentionBackendEnum.SAGE_ATTN,
             AttentionBackendEnum.AITER,
             AttentionBackendEnum.TORCH_SDPA,
         }

python/sglang/multimodal_gen/configs/pipeline_configs/minimax_h3.py

@@ -169,12 +169,6 @@ class MiniMaxH3PipelineConfig(PipelineConfig):
     def validate_server_args(self, server_args) -> None:
         # Reject known-inexact VAE modes before any large component download.
         self.vae_config.resolved_parallel_decode_mode()
-        attention_backend = self._server_arg_value(server_args.attention_backend)
-        if str(attention_backend).strip().lower() == "sage_attn":
-            raise ValueError(
-                "MiniMax-H3 does not support SageAttention: the current packed "
-                "varlen path does not preserve model output"
-            )
 
     def select_vae_weight_files(
         self,

python/sglang/multimodal_gen/runtime/layers/attention/backends/sage_attn.py

@@ -17,6 +17,21 @@ from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
 logger = init_logger(__name__)
 
 
+def _trailing_padding_used_len(
+    *,
+    total_tokens: int,
+    max_seqlen: int,
+    bounds: tuple[int, ...],
+) -> int | None:
+    """Return live token count for H3-style [0, used, total] trailing padding."""
+    if len(bounds) != 3:
+        return None
+    start, used, total = bounds
+    if start != 0 or used >= total or total != total_tokens or used != max_seqlen:
+        return None
+    return used
+
+
 class SageAttentionBackend(AttentionBackend):
     accept_output_buffer: bool = True
 
@@ -72,3 +87,68 @@ class SageAttentionImpl(AttentionImpl):
             output, softmax_lse = output
             return output, softmax_lse
         return output
+
+    def forward_varlen(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        *,
+        cu_seqlens: torch.Tensor,
+        max_seqlen: int,
+        cu_seqlens_host: tuple[int, ...] | None = None,
+    ) -> torch.Tensor:
+        bounds = (
+            cu_seqlens_host
+            if cu_seqlens_host is not None
+            else tuple(int(x) for x in cu_seqlens.tolist())
+        )
+        return self._sage_packed(
+            query.contiguous(),
+            key.contiguous(),
+            value.contiguous(),
+            bounds=bounds,
+            max_seqlen=max_seqlen,
+        )
+
+    def _sage_packed(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        *,
+        bounds: tuple[int, ...],
+        max_seqlen: int,
+    ) -> torch.Tensor:
+        # MiniMax-H3 packs one live document as bounds=(0, used, total):
+        # [0, used) are real tokens; [used, total) is 64-aligned tail padding.
+        used = _trailing_padding_used_len(
+            total_tokens=query.shape[0],
+            max_seqlen=max_seqlen,
+            bounds=bounds,
+        )
+        if used is not None:
+            live_out = self.forward(
+                query[:used].unsqueeze(0),
+                key[:used].unsqueeze(0),
+                value[:used].unsqueeze(0),
+                None,
+            )[0]
+            if used == query.shape[0]:
+                return live_out
+            # Keep padded tail at zero so downstream masked rows stay inactive.
+            output = torch.zeros_like(query)
+            output[:used] = live_out
+            return output
+
+        output = torch.empty_like(query)
+        for start, stop in zip(bounds[:-1], bounds[1:]):
+            if start == stop:
+                continue
+            output[start:stop] = self.forward(
+                query[start:stop].unsqueeze(0),
+                key[start:stop].unsqueeze(0),
+                value[start:stop].unsqueeze(0),
+                None,
+            )[0]
+        return output

python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/release_metadata.py

@@ -153,12 +153,6 @@ class MiniMaxH3PartitionAdmissionStage(PipelineStage):
                 f"quality must be one of {list(QUALITY_LEVELS)}, got {quality!r}"
             )
         high_quality = quality == "high"
-        attention_backend = str(server_args.attention_backend or "").strip().lower()
-        if attention_backend == "sage_attn" and not batch.is_warmup:
-            raise ValueError(
-                "MiniMax-H3 does not support SageAttention: the current packed "
-                "varlen path does not preserve model output"
-            )
         if high_quality and not batch.is_warmup:
             server_args.pipeline_config.validate_quality_deployment(server_args)
             plan = minimax_h3_plan_from_batch(batch)

python/sglang/multimodal_gen/test/unit/test_minimax_h3_admission.py

@@ -301,8 +301,7 @@ def test_quality_admission_fails_closed_outside_validated_request():
     batch.sampling_params.quality = "lossless"
     batch.num_inference_steps = 50
     server_args.attention_backend = "sage_attn"
-    with pytest.raises(ValueError, match="does not support SageAttention"):
+    assert stage.forward(batch, server_args) is batch
-        stage.forward(batch, server_args)
 
     batch.sampling_params.quality = "ultra"
     server_args.attention_backend = None