Test that attention gets the real max_length; state what FlashAttention-4 needs

The actor test's pad segment was as long as its longest document, so it could
not tell get_batch's max_seqlen from the real one. The docs now list every
architecture that passes the boundaries and say the rest cannot use
flash_attention_4 yet.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Author: HJSang

docs/user-guide/training-backend.md

@@ -381,8 +381,9 @@ targets are Megatron-only: both `--offload-train-target disk` and
 - `--attn-implementation` picks the `transformers` attention backend: `flash_attention_2` by
   default, with `flash_attention_3`, `sdpa` and `eager` passed straight through.
   `flash_attention_4` uses the FlashAttention-4 that SGLang installs, so it trains without a
-  `flash-attn` build. It needs an adapter that passes the packed-document boundaries: Qwen3,
-  Qwen3-MoE, Qwen3-VL, GLM-4.7-Flash and Qwen3.5.
+  `flash-attn` build. It needs the packed-document boundaries from the architecture's adapter:
+  Qwen3, Qwen3-MoE, Qwen3-VL, GLM-4.7-Flash, Qwen3.5, Qwen3.5-MoE and Qwen3-Next pass them; other
+  architectures (no adapter, or remote code) do not and cannot use it yet.
 - An architecture with fussier numerics can register its own policy, see
   [when an HF model needs help](#going-deeper-when-an-hf-model-needs-help).
 

miles/backends/fsdp_utils/adaptations/arch_adapter.py

@@ -26,7 +26,7 @@ class ArchAdapter:
     verified: bool = False
     # None means the architecture has no MoE router to replay.
     routing_replay: RoutingReplayAdapter | None = None
-    # stock HF attention reads the varlen boundaries; remote code may not accept them
+    # stock HF modeling takes the packed boundaries (FlashAttention-4 needs them); remote code may not
     passes_attention_packing_kwargs: bool = False
 
     def resolve_precision(self, base: PrecisionPolicy, args: Namespace) -> PrecisionPolicy:

tests/fast/backends/fsdp_utils/test_actor.py

@@ -91,21 +91,22 @@ def test_packing_kwargs_follow_the_batch_boundaries_so_padding_stays_one_segment
         """Pad tokens carry position 0; boundaries derived from `position_ids` would make each one a document."""
         actor = object.__new__(actor_module.FSDPTrainRayActor)
         actor.arch_adapter = resolve_arch_adapter(SimpleNamespace(model_type="qwen3_5_text"))
-        cu_seqlens_host = (0, 3, 5, 8)  # two documents, then the 3-token pad segment get_batch appends
+        cu_seqlens_host = (0, 3, 5, 9)  # two documents, then the 4-token pad segment get_batch appends
         batch = {
-            "tokens": torch.zeros(1, 8, dtype=torch.long),
-            "position_ids": torch.tensor([[0, 1, 2, 0, 1, 0, 0, 0]]),
+            "tokens": torch.zeros(1, 9, dtype=torch.long),
+            "position_ids": torch.tensor([[0, 1, 2, 0, 1, 0, 0, 0, 0]]),
             "cu_seqlens": torch.tensor(cu_seqlens_host, dtype=torch.int32),
             "cu_seqlens_host": cu_seqlens_host,
-            "max_seqlen": 3,
+            "max_seqlen": 4,  # get_batch's, which counts the pad segment
             "unconcat_tokens": [torch.zeros(3, dtype=torch.long), torch.zeros(2, dtype=torch.long)],
         }
 
         model_args = actor._get_model_inputs_args(batch)
 
         assert model_args["cu_seq_lens_q"] is batch["cu_seqlens"]
+        # the longest real document: the pad segment changes FlashAttention's launch plan
         assert model_args["max_length_q"] == 3
-        assert model_args["seq_idx"].tolist() == [[0, 0, 0, 1, 1, 2, 2, 2]]
+        assert model_args["seq_idx"].tolist() == [[0, 0, 0, 1, 1, 2, 2, 2, 2]]
 
 
 class TestFSDPTrainParallelConfigWiring: