Add GLM-4.6 nightly benchmark coverage

.github/workflows/nightly-benchmark.yml

@@ -233,7 +233,6 @@ jobs:
           - { id: meta-llama/Llama-3.1-8B-Instruct, slug: meta-llama-Llama-3.1-8B-Instruct, test_class: TestNightlyLlama8bMulti }
           - { id: Qwen/Qwen2.5-7B-Instruct, slug: Qwen-Qwen2.5-7B-Instruct, test_class: TestNightlyQwen7bMulti }
           - { id: Qwen/Qwen3-30B-A3B, slug: Qwen-Qwen3-30B-A3B, test_class: TestNightlyQwen30bMulti }
-          - { id: openai/gpt-oss-20b, slug: openai-gpt-oss-20b, test_class: TestNightlyGptOss20bMulti }
         variant:
           - { id: sglang, runtime: sglang, grpc_only: "false", setup_vllm: false, setup_trtllm: false, extra_deps: "genai-bench" }
           - { id: vllm,   runtime: vllm,   grpc_only: "false", setup_vllm: true, setup_trtllm: false, extra_deps: "genai-bench" }
@@ -347,6 +346,9 @@ jobs:
           - { id: meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8, slug: meta-llama-Llama-4-Maverick-17B-128E-Instruct-FP8, test_class: TestNightlyLlama4MaverickSingle }
           # Pausing MiniMax M2 nightly benchmark
           # - { id: minimaxai/minimax-m2, slug: minimaxai-minimax-m2, test_class: TestNightlyMinimaxM2Single }
+          - id: zai-org/GLM-4.6
+            slug: zai-org-GLM-4.6
+            test_class: TestNightlyGlm46Single
         variant:
           - { id: sglang, runtime: sglang, grpc_only: "false", setup_vllm: false, setup_trtllm: false, extra_deps: "genai-bench" }
           - { id: vllm,   runtime: vllm,   grpc_only: "false", setup_vllm: true, setup_trtllm: false, extra_deps: "genai-bench" }

e2e_test/benchmarks/test_nightly_perf.py

@@ -4,10 +4,11 @@
 and concurrency levels. No performance thresholds — results are uploaded
 as artifacts for tracking over time.
 
-Each model has Single (1 worker) and Multi (N workers) classes, both
-parametrized with http and grpc backends. The workflow matrix crosses
-model × variant (single/multi × sglang/vllm); both protocols run for
-all runtimes.
+Each model always has a Single (1 worker) class, and models with
+`multi_workers > 1` also get a Multi (N workers) class. Both variants
+are parametrized with http and grpc backends. The workflow matrix
+crosses the generated variants with sglang and vllm runtimes, and both
+protocols run for all runtimes.
 
 genai-bench defaults (omitted flags):
   - Concurrency: [1, 2, 4, 8, 16, 32, 64, 128, 256]
@@ -102,6 +103,7 @@ def _run_nightly(setup_backend, genai_bench_runner, model_id, worker_count=1, **
     ("Qwen/Qwen3-30B-A3B", "Qwen30b", 4, ["http", "grpc"], {}),
     ("openai/gpt-oss-20b", "GptOss20b", 1, ["http", "grpc"], {}),
     ("minimaxai/minimax-m2", "MinimaxM2", 1, ["http", "grpc"], {}),
+    ("zai-org/GLM-4.6", "Glm46", 1, ["http", "grpc"], {}),
     (
         "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
         "Llama4Maverick",
@@ -142,12 +144,16 @@ def test_nightly_perf(self, setup_backend, genai_bench_runner):
 
 
 for _model_id, _name, _multi_workers, _backends, _extra in _NIGHTLY_MODELS:
-    for _suffix, _count in [("Single", 1), ("Multi", _multi_workers)]:
+    variants = [("Single", 1)]
+    if _multi_workers > 1:
+        variants.append(("Multi", _multi_workers))
+
+    for _suffix, _count in variants:
         _cls_name = f"TestNightly{_name}{_suffix}"
         _cls = _make_test_class(_model_id, _count, _backends, _extra)
         _cls.__name__ = _cls_name
         _cls.__qualname__ = _cls_name
         globals()[_cls_name] = _cls
 
 # Clean up loop variables from module namespace
-del _model_id, _name, _multi_workers, _backends, _extra, _suffix, _count, _cls_name, _cls
+del _model_id, _name, _multi_workers, _backends, _extra, _suffix, _count, _cls_name, _cls, variants

e2e_test/infra/model_specs.py

@@ -96,6 +96,14 @@ def _resolve_model_path(hf_path: str) -> str:
         "worker_args": ["--trust-remote-code"],
         "vllm_args": ["--trust-remote-code"],
     },
+    # GLM-4.6 - nightly benchmarks
+    "zai-org/GLM-4.6": {
+        "model": _resolve_model_path("zai-org/GLM-4.6"),
+        "tp": 8,
+        "features": ["chat", "streaming", "function_calling", "reasoning"],
+        "worker_args": ["--trust-remote-code"],
+        "vllm_args": ["--trust-remote-code"],
+    },
     # Vision-language model for multimodal benchmarks (MMMU)
     "Qwen/Qwen3-VL-8B-Instruct": {
         "model": _resolve_model_path("Qwen/Qwen3-VL-8B-Instruct"),