support asymmetric pd-mtp via mock spec hidden

python/sglang/srt/managers/scheduler.py

@@ -892,18 +892,12 @@ def init_disaggregation(self):
             self.req_to_metadata_buffer_idx_allocator = ReqToMetadataIdxAllocator(
                 buffer_size
             )
+            # Full-size buffer on both sides so the wire layout aligns
+            # under asymmetric P/D where one side may not run spec.
             self.disagg_metadata_buffers = MetadataBuffers(
                 buffer_size,
-                hidden_size=(
-                    model_config.spec_hidden_size
-                    if self.spec_algorithm.is_eagle()
-                    else 16  # minimal padding size for RDMA
-                ),
-                hidden_states_dtype=(
-                    model_config.dtype
-                    if self.spec_algorithm.is_eagle()
-                    else torch.float32
-                ),
+                hidden_size=model_config.spec_hidden_size,
+                hidden_states_dtype=model_config.dtype,
                 custom_mem_pool=self.token_to_kv_pool_allocator.get_kvcache().maybe_get_custom_mem_pool(),
             )
 
@@ -946,20 +940,12 @@ def init_disaggregation(self):
             self.req_to_metadata_buffer_idx_allocator = ReqToMetadataIdxAllocator(
                 buffer_size
             )
+            # See decode branch above. Asymmetric P/D: prefill without a
+            # spec module ships zeros, decode mocks first-step conditioning.
             self.disagg_metadata_buffers = MetadataBuffers(
                 buffer_size,
-                hidden_size=(
-                    model_config.spec_hidden_size
-                    if self.spec_algorithm.is_eagle()
-                    or self.spec_algorithm.is_standalone()
-                    else 16  # minimal padding size for RDMA
-                ),
-                hidden_states_dtype=(
-                    model_config.dtype
-                    if self.spec_algorithm.is_eagle()
-                    or self.spec_algorithm.is_standalone()
-                    else torch.float32
-                ),
+                hidden_size=model_config.spec_hidden_size,
+                hidden_states_dtype=model_config.dtype,
                 custom_mem_pool=self.token_to_kv_pool_allocator.get_kvcache().maybe_get_custom_mem_pool(),
             )
 

test/registered/8-gpu-models/test_dsv4_pd_disagg_nixl.py

@@ -1,22 +1,7 @@
-"""DSv4 Flash PD-disaggregation test with NIXL transfer backend.
-
-Topology (1 H200 node, 8 GPUs total):
-  - Prefill: GPU 0-3, tp=4 — pure TP, **no EP** (no deepep), no DP
-    attention. Optimized for throughput on long prompts; each rank
-    holds the full MoE weights, no all-to-all dispatch traffic.
-    Spec config matches decode (PD ferry currently assumes symmetric
-    spec on both sides) so the prefill -> decode metadata buffer
-    is sized correctly for the spec module's hidden shape.
-  - Decode:  GPU 4-7, tp=4 dp=4 enable-dp-attention + deepep + EAGLE
-    MTP — optimized for low-latency decode with spec decoding and
-    expert parallelism.
-  - Mini load balancer fronting both.
-
-Both sides use DSv4 Flash FP8 weights. Transfer backend is NIXL
-(the focus of recent nixl/conn.py forward-delta work; this test is
-the e2e check that the generic `send_state` / shared buffer-pool
-changes do not break PD).
-"""
+"""DSv4 Flash PD-disagg with NIXL backend, asymmetric: prefill is pure
+TP with no spec module, decode runs EAGLE MTP. Prefill ships a zero-
+init hidden buffer; decode mocks first-step conditioning, verify keeps
+the output correct."""
 
 import unittest
 from types import SimpleNamespace
@@ -38,9 +23,7 @@
 
 DSV4_FLASH_ENV = {
     "SGLANG_DSV4_FP4_EXPERTS": "0",
-    # Decode side runs MTP with num_draft_tokens=4 → dispatch input scales
-    # by ~4x, so default 256 overflows once cuda-graph-max-bs * draft > 256.
-    # 1024 covers bs=128 * 4 with headroom (no-op on prefill which has no EP).
+    # MTP num_draft_tokens=4 scales dispatch by ~4x; 256 overflows at bs=128.
     "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "1024",
     "SGLANG_JIT_DEEPGEMM_PRECOMPILE": "0",
 }
@@ -65,9 +48,6 @@ def setUpClass(cls):
 
     @classmethod
     def start_prefill(cls):
-        # Prefill: TP=4 (no EP, no DP attention). EAGLE config mirrors decode
-        # so the metadata buffer is sized for the spec module's hidden shape;
-        # PD ferry currently assumes both sides agree on the spec algorithm.
         prefill_args = [
             "--trust-remote-code",
             "--disaggregation-mode",
@@ -86,14 +66,6 @@ def start_prefill(cls):
             "4",
             "--disaggregation-decode-dp",
             "4",
-            "--speculative-algorithm",
-            "EAGLE",
-            "--speculative-num-steps",
-            "3",
-            "--speculative-eagle-topk",
-            "1",
-            "--speculative-num-draft-tokens",
-            "4",
             *cls.transfer_backend,
             *cls.rdma_devices,
         ]
@@ -107,7 +79,6 @@ def start_prefill(cls):
 
     @classmethod
     def start_decode(cls):
-        # Decode: TP=4 + DP=4 attention + deepep EP + EAGLE MTP.
         decode_args = [
             "--trust-remote-code",
             "--disaggregation-mode",