[Bugfix] Stamp capture-time num_tokens_per_req in multi-layer EAGLE; close jit_kernel CI filter gaps (#31367)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>

.github/workflows/_pr-test-check-changes.yml

@@ -103,6 +103,9 @@ jobs:
               - "python/pyproject.toml"
               - "python/sglang/jit_kernel/**"
               - "test/registered/jit/**"
+              # sglang.kernels is the migrated kernel namespace (RFC #29630 / #30044); the
+              # base-b-kernel suites import it directly, so kernel edits must run them.
+              - "python/sglang/kernels/**"
             sgl_kernel:
               # Intentionally excludes ".github/workflows/pr-test-sgl-kernel.yml" —
               # see API-side detector below for rationale.

python/sglang/srt/speculative/multi_layer_eagle_draft_extend_cuda_graph_runner.py

@@ -273,6 +273,7 @@ class MultiLayerEagleDraftExtendCudaGraphRunner(DecodeCudaGraphRunner):
             num_correct_drafts=num_correct_drafts,
             num_accept_tokens=num_accept_tokens,
         )
+        spec_info.num_tokens_per_req = self.num_tokens_per_req
         spec_info.positions = None
 
         capture_mode = (

test/registered/attention/test_kda_decode_flashinfer.py

@@ -10,11 +10,11 @@ import torch
 
 from sglang.test.ci.ci_register import register_cuda_ci
 
-# SM100 single-GPU kernel-unit suite, same slot as the CuteDSL KDA prefill test.
+# SM100 suite, same slot as the CuteDSL KDA prefill test.
 # Disabled in public CI until the B200 runner image ships recurrent_kda.
 register_cuda_ci(
     est_time=60,
-    stage="base-b-kernel-unit",
+    stage="base-b",
     runner_config="4-gpu-b200",
     disabled="recurrent_kda (SM100 KDA decode) not guaranteed in public CI FlashInfer build",
 )

test/registered/attention/test_trtllm_mha_graph_metadata.py

@@ -23,7 +23,7 @@ from sglang.srt.model_executor.forward_batch_info import ForwardMode
 from sglang.test.ci.ci_register import register_cuda_ci
 
 # trtllm_mha kernels are sm100-only; run this kernel-unit test on Blackwell.
-register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=30, stage="base-b", runner_config="4-gpu-b200")
 
 DEVICE = "cuda"
 PAGE_SIZE = 128

test/registered/jit/test_custom_all_reduce.py

@@ -45,7 +45,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 
 register_cuda_ci(
     est_time=300,
-    stage="base-b-kernel-unit",
+    stage="extra-b",
     runner_config="8-gpu-h200",
 )
 # Nightly is not redundant here: it sets SGLANG_JIT_KERNEL_RUN_FULL_TESTS=1 to expand get_ci_test_range sweeps.

test/registered/jit/test_dsv32_indexer_fusion.py

@@ -31,7 +31,7 @@ from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
 _is_hip = is_hip()
 
-register_cuda_ci(est_time=45, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_cuda_ci(est_time=45, stage="base-b", runner_config="1-gpu-large")
 register_amd_ci(est_time=45, suite="jit-kernel-unit-test-amd")
 
 HEAD_DIM = 128

test/registered/jit/test_hicache.py

@@ -14,7 +14,7 @@ from sglang.srt.mem_cache.pool_host.mla import MLATokenToKVPoolHost
 from sglang.srt.utils import is_cuda, is_hip, is_npu, is_xpu
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=10, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_cuda_ci(est_time=10, stage="base-b", runner_config="1-gpu-large")
 
 pytestmark = pytest.mark.skipif(
     not torch.cuda.is_available()

test/registered/jit/test_hicache_page_first_write_back.py

@@ -24,7 +24,7 @@ from sglang.srt.mem_cache.pool_host.mla import MLATokenToKVPoolHost
 from sglang.srt.utils import is_cuda, is_hip, is_npu, is_xpu
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=15, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_cuda_ci(est_time=15, stage="base-b", runner_config="1-gpu-large")
 register_amd_ci(est_time=30, stage="jit-kernel-unit", runner_config="amd")
 
 pytestmark = pytest.mark.skipif(

test/registered/jit/test_kvcacheio_asymmetric.py

@@ -7,7 +7,7 @@ import torch
 from sglang.srt.mem_cache.pool_host.mha import AsymmetricMHATokenToKVPoolHost
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 
-register_cuda_ci(est_time=10, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_cuda_ci(est_time=10, stage="base-b", runner_config="1-gpu-large")
 register_amd_ci(est_time=10, suite="nightly-amd-kernel-1-gpu", nightly=True)
 
 # These tests use AsymmetricMHATokenToKVPoolHost methods and let that class call

test/registered/jit/test_moe_wna16_marlin.py

@@ -20,7 +20,7 @@ from sglang.test.test_marlin_utils import (
     marlin_quantize,
 )
 
-register_cuda_ci(est_time=10, stage="base-b-kernel-unit", runner_config="1-gpu-large")
+register_cuda_ci(est_time=10, stage="base-b", runner_config="1-gpu-large")
 
 
 def _has_aot_moe_wna16_marlin_gemm() -> bool:

test/registered/jit/test_symm_mem_all_gather.py

@@ -34,7 +34,7 @@ from sglang.srt.distributed.device_communicators.triton_symm_mem_ag import (
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=240, stage="base-b-kernel-unit", runner_config="8-gpu-h200")
+register_cuda_ci(est_time=240, stage="extra-b", runner_config="8-gpu-h200")
 # Nightly is not redundant here: it sets SGLANG_JIT_KERNEL_RUN_FULL_TESTS=1 to expand get_ci_test_range sweeps.
 register_cuda_ci(est_time=240, suite="nightly-kernel-8-gpu-h200", nightly=True)
 

test/registered/jit/test_tp_qknorm.py

@@ -29,7 +29,7 @@ from sglang.test.ci.ci_register import register_cuda_ci
 
 register_cuda_ci(
     est_time=300,
-    stage="base-b-kernel-unit",
+    stage="extra-b",
     runner_config="8-gpu-h200",
 )