Install DeepEP from release wheels (#33932)

test/registered/attention/unittests/dense/test_tbo.py

@@ -102,7 +102,7 @@ class TestTboAttnDenseAttentionBackendCorrectness(CustomTestCase):
         only by the in_capture=True branch (via ``_bind_metadata_buffers``).
         If TBO short-circuits its capture to its own replay path, those dicts
         are empty and replay raises ``KeyError: bs``. Reproduces the
-        deepep-4-gpu-h100 failure where
+        4-gpu-h100 failure where
         ``flashattention_backend.target_verify_metadata[bs]`` lookup blew up
         during ``init_device_graphs``.
 

test/registered/cp/test_deepseek_v3_cp_single_node.py

@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=500, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=500, stage="extra-b", runner_config="8-gpu-h200")
 
 DEEPSEEK_V3_MODEL_PATH = "deepseek-ai/DeepSeek-V3-0324"
 

test/registered/cp/test_deepseek_v4_flash_fp4_b200_cp.py

@@ -22,7 +22,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=235, stage="extra-b", runner_config="deepep-4-gpu-b200")
+register_cuda_ci(est_time=235, stage="extra-b", runner_config="4-gpu-b200")
 
 MODEL = "deepseek-ai/DeepSeek-V4-Flash"
 SERVER_LAUNCH_TIMEOUT = 3600

test/registered/cp/test_dsa_prefill_cp.py

@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=320, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=320, stage="extra-b", runner_config="8-gpu-h200")
 GLM52_MODEL_PATH = "zai-org/GLM-5.2-FP8"
 SERVER_LAUNCH_TIMEOUT = max(DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, 1800)
 

test/registered/cp/test_dsa_prefill_cp_legacy.py

@@ -13,7 +13,7 @@ from sglang.test.test_utils import (
     write_github_step_summary,
 )
 
-register_cuda_ci(est_time=616, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=616, stage="extra-b", runner_config="8-gpu-h200")
 DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 
 

test/registered/cp/test_gqa_prefill_cp.py

@@ -11,7 +11,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=500, stage="extra-b", runner_config="deepep-4-gpu-h100")
+register_cuda_ci(est_time=500, stage="extra-b", runner_config="4-gpu-h100")
 
 GQA_MODEL_PATH = "Qwen/Qwen3-30B-A3B-FP8"
 

test/registered/disaggregation/test_disaggregation_dsv4.py

@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=500, stage="base-c", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=500, stage="base-c", runner_config="8-gpu-h200")
 
 DSV4_FLASH_MODEL = "sgl-project/DeepSeek-V4-Flash-FP8"
 

test/registered/disaggregation/test_disaggregation_hisparse.py

@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=1000, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=1000, stage="extra-b", runner_config="8-gpu-h200")
 
 DSV4_FLASH_MODEL = "sgl-project/DeepSeek-V4-Flash-FP8"
 DSV4_FLASH_LOADER_CONFIG = '{"enable_multithread_load": true, "num_threads": 64}'

test/registered/ep/test_deepep_large.py

@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=528, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=528, stage="extra-b", runner_config="8-gpu-h200")
 
 DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
 

test/registered/ep/test_deepep_small.py

@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=478, stage="base-c", runner_config="deepep-4-gpu-h100")
+register_cuda_ci(est_time=478, stage="base-c", runner_config="4-gpu-h100")
 
 
 class TestPureDP(CustomTestCase):

test/registered/ep/test_mooncake_ep_small.py

@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
 register_cuda_ci(
     est_time=189,
     stage="base-c",
-    runner_config="deepep-4-gpu-h100",
+    runner_config="4-gpu-h100",
     disabled="Temporarily disabled until the next Mooncake release includes the PyTorch 2.13 collective forwarding fix.",
 )
 

test/registered/ep/test_tbo_shared_experts_fusion.py

@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=900, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=900, stage="extra-b", runner_config="8-gpu-h200")
 
 DEEPSEEK_V3_MODEL_PATH = "deepseek-ai/DeepSeek-V3-0324"
 

test/registered/models_e2e/test_deepseek_v4_flash_fp4_b200.py

@@ -4,7 +4,7 @@ Launches TP=4 with flashinfer_mxfp4 MoE runner + EAGLE speculative decoding.
 Runs 12 ServerSanity probes (correctness, streaming, concurrency, determinism)
 plus a GSM8K accuracy gate.
 
-Registry: base-c-test-deepep-4-gpu-b200 (per-commit, 4x B200)
+Registry: base-c-test-4-gpu-b200 (per-commit, 4x B200)
 """
 
 import unittest
@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=465, stage="base-c", runner_config="deepep-4-gpu-b200")
+register_cuda_ci(est_time=465, stage="base-c", runner_config="4-gpu-b200")
 
 MODEL = "deepseek-ai/DeepSeek-V4-Flash"
 SERVER_LAUNCH_TIMEOUT = 3600

test/registered/models_e2e/test_deepseek_v4_flash_fp4_h200.py

@@ -7,7 +7,7 @@ plus a GSM8K accuracy gate.
 Also covers SGLANG_DSV4_FP4_DEQUANT=1 (TP=8): FP4 experts dequantized to FP8
 during loading and served through the plain FP8 MoE path.
 
-Registry: base-c-test-deepep-8-gpu-h200 (per-commit, 8x H200)
+Registry: base-c-test-8-gpu-h200 (per-commit, 8x H200)
 """
 
 import unittest
@@ -24,7 +24,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=600, stage="base-c", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=600, stage="base-c", runner_config="8-gpu-h200")
 
 
 def _flashinfer_has_sm90_cutlass_mxfp4() -> bool:

test/registered/models_e2e/test_deepseek_v4_flash_fp4_megamoe_b200.py

@@ -4,7 +4,7 @@ Launches TP=4 with flashinfer_mxfp4 MoE runner + EAGLE speculative decoding.
 Runs 12 ServerSanity probes (correctness, streaming, concurrency, determinism)
 plus a GSM8K accuracy gate.
 
-Registry: extra-b-test-deepep-4-gpu-b200 (label-gated, 4x B200)
+Registry: extra-b-test-4-gpu-b200 (label-gated, 4x B200)
 """
 
 import unittest
@@ -21,7 +21,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=900, stage="extra-b", runner_config="deepep-4-gpu-b200")
+register_cuda_ci(est_time=900, stage="extra-b", runner_config="4-gpu-b200")
 
 MODEL = "deepseek-ai/DeepSeek-V4-Flash"
 SERVER_LAUNCH_TIMEOUT = 3600

test/registered/models_e2e/test_deepseek_v4_flash_fp8_h200.py

@@ -5,7 +5,7 @@ with FP4 experts disabled via SGLANG_DSV4_FP4_EXPERTS=0.
 Runs 12 ServerSanity probes (correctness, streaming, concurrency, determinism)
 plus a GSM8K accuracy gate.
 
-Registry: extra-b-test-deepep-8-gpu-h200 (label-gated, 8x H200 — only 4 used by TP=4)
+Registry: extra-b-test-8-gpu-h200 (label-gated, 8x H200 — only 4 used by TP=4)
 """
 
 import unittest
@@ -22,7 +22,7 @@ from sglang.test.test_utils import (
     try_cached_model,
 )
 
-register_cuda_ci(est_time=560, stage="extra-b", runner_config="deepep-8-gpu-h200")
+register_cuda_ci(est_time=560, stage="extra-b", runner_config="8-gpu-h200")
 
 MODEL_FP8 = "sgl-project/DeepSeek-V4-Flash-FP8"
 SERVER_LAUNCH_TIMEOUT = 3600

test/registered/rl/test_return_routed_experts.py

@@ -24,7 +24,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=400, stage="extra-b", runner_config="deepep-4-gpu-h100")
+register_cuda_ci(est_time=400, stage="extra-b", runner_config="4-gpu-h100")
 
 # FP8 variant of Qwen3-30B-A3B: required because DeepEP normal/LL fast paths in
 # ep_moe/layer.py only run for {Fp8Config (via deep_gemm), W4AFp8Config, aiter,

test/run_suite.py

@@ -83,9 +83,6 @@ PER_COMMIT_SUITES = {
         "base-c-test-8-gpu-h200",
         "base-c-test-8-gpu-b200",
         "base-c-test-8-gpu-b300",
-        "base-c-test-deepep-4-gpu-h100",
-        "base-c-test-deepep-4-gpu-b200",
-        "base-c-test-deepep-8-gpu-h200",
         # extra-a / extra-b: label-gated PR opt-in suites in pr-test-extra.yml
         # (tests still tagged per-commit but skipped on default PR runs).
         "extra-a-test-1-gpu-small",
@@ -94,9 +91,6 @@ PER_COMMIT_SUITES = {
         "extra-b-test-4-gpu-h100",
         "extra-b-test-4-gpu-b200",
         "extra-b-test-8-gpu-h200",
-        "extra-b-test-deepep-4-gpu-h100",
-        "extra-b-test-deepep-4-gpu-b200",
-        "extra-b-test-deepep-8-gpu-h200",
     ],
     HWBackend.NPU: [
         "base-a-test-1-gpu-small",