patches: glm53-dflash-01 v3 -- declare supports_eagle3 as a class attribute

v2 (inheriting SupportsEagle3 as a base) taught the real lesson: the isinstance
check PASSES and the class then dies at creation with
'Cannot create a consistent method resolution order (MRO) for bases ...'
-- GLM's outer wrapper already inherits vLLM Protocols (SupportsPP,
MixtureOfExperts, HasInnerState, IsHybrid) whose ordering is incompatible with
appending another Protocol base.

v3 avoids the MRO entirely: declare
    supports_eagle3: ClassVar[Literal[True]] = True
as a class attribute on Glm5NextForCausalLM.  Runtime-checkable Protocols do
presence-based isinstance, so the bare flag satisfies supports_eagle3(model)
without touching the MRO.  The aux-layer plumbing stays on the trunk via
EagleModelMixin.

base 008eeaa2... -> patched 4c9a632b..., apply-verified against the pinned
image (patch -p1 reproduces the sha256 and matches the mounted dflash-patch
file byte for byte).

scripts/patches/glm53-dflash-01-aux-hidden-states.patch

@@ -18,12 +18,19 @@
 #
 # THE CONTRACT (read out of this image, not assumed):
 #   * interfaces.py:1663 -- `supports_eagle3: ClassVar[Literal[True]] = True`.  THE FLAG
-#     IS A REQUIRED MEMBER FOR `isinstance`: satisfying the Protocol with methods alone
+#     IS A REQUIRED MEMBER for runtime-checkable presence-based isinstance, so v1
-#     still raises the same RuntimeError -- v1 of this patch did exactly that and the boot
+#     (methods only) failed identically.  BUT v2 (inheriting the Protocol) taught the
-#     failed identically.  INHERIT the Protocol instead.  That is the idiom in this image
+#     sharper lesson: adding SupportsEagle3 to Glm5NextForCausalLM's bases passes
-#     (LagunaForCausalLM, MiMoV2FlashForCausalLM and DFlashLagunaForCausalLM all list
+#     isinstance and then dies at class creation with
-#     SupportsEagle3 among their bases) and it brings the flag, SupportsEagleBase's
+#     "Cannot create a consistent method resolution order (MRO) for bases ..."
-#     has_own_lm_head/has_own_embed_tokens, AND working non-stub defaults for both methods.
+#     -- because GLM's outer wrapper already inherits vLLM Protocols that are NOT
+#     Protocol-derived mixins in a compatible order.  The fix that avoids the MRO
+#     entirely: declare the flag as a class attribute on the wrapper:
+#         supports_eagle3: ClassVar[Literal[True]] = True
+#     which satisfies presence-based isinstance without touching the MRO, and pairs
+#     with the EagleModelMixin on the trunk that supplies the real aux-layer plumbing.
+#     (SupportsEagleBase's has_own_lm_head/has_own_embed_tokens default to False, which
+#     is what a wrapper that delegates embed/lm_head to its trunk wants.)
 #   * interfaces.py:1673-1719 -- the inherited defaults unwrap
 #     `holder = getattr(parent_ref, "model", parent_ref)`, assert
 #     `isinstance(holder, EagleModelMixin)`, set `holder._set_aux_hidden_state_layers`, and
@@ -50,12 +57,12 @@
 #
 # TARGET: vllm/models/glm5next/common/model.py   (inside the pinned image)
 #   base    sha256 008eeaa2a07e091465319f86900b5a321d921238723ccbe22e6b9e971e1d2b28
-#   patched sha256 5eea89c51d4ab37ef823526286b5103e39996802ca08aa1934f6373fc4728062
+#   patched sha256 4c9a632b5173fdefe0672b810b54f91b88ed65decf0f4a76fde92527970bee8e
 #
 # APPLY: `patch -p1` from the vllm package root, or bind-mount the patched file over
 # site-packages (which is what a scratch-launcher DFLASH_PATCH=1 does).
---- a/vllm/models/glm5next/common/model.py
+--- orig/vllm/models/glm5next/common/model.py	2026-09-30 15:57:08.910029614 -0700
-+++ b/vllm/models/glm5next/common/model.py
++++ new/vllm/models/glm5next/common/model.py	2026-09-30 17:04:20.104521484 -0700
 @@ -95,6 +95,12 @@
  logger = init_logger(__name__)
  
@@ -103,17 +110,12 @@
          return hidden_states
  
      def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
-@@ -935,7 +950,12 @@
+@@ -937,6 +952,8 @@
- 
- 
  class Glm5NextForCausalLM(
--    nn.Module, HasInnerState, SupportsPP, MixtureOfExperts, IsHybrid
+     nn.Module, HasInnerState, SupportsPP, MixtureOfExperts, IsHybrid
-+    nn.Module,
-+    HasInnerState,
-+    SupportsPP,
-+    MixtureOfExperts,
-+    IsHybrid,
-+    SupportsEagle3,
  ):
++    supports_eagle3: ClassVar[Literal[True]] = True
++
      def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
          super().__init__()
+         self.model_config = vllm_config.model_config