[None][test] Prune Llama-3.1 8b model from tests

tests/integration/defs/conftest.py

@@ -680,18 +680,6 @@ def llama_model_root(request):
     if request.param == "TinyLlama-1.1B-Chat-v1.0":
         llama_model_root = os.path.join(models_root, "llama-models-v2",
                                         "TinyLlama-1.1B-Chat-v1.0")
-    elif request.param == "llama-3.1-8b":
-        llama_model_root = os.path.join(models_root, "llama-3.1-model",
-                                        "Meta-Llama-3.1-8B")
-    elif request.param == "llama-3.1-8b-instruct-hf-fp8":
-        llama_model_root = os.path.join(models_root, "llama-3.1-model",
-                                        "Llama-3.1-8B-Instruct-FP8")
-    elif request.param == "llama-3.1-8b-instruct":
-        llama_model_root = os.path.join(models_root, "llama-3.1-model",
-                                        "Llama-3.1-8B-Instruct")
-    elif request.param == "llama-3.1-8b-hf-nvfp4":
-        llama_model_root = os.path.join(models_root, "nvfp4-quantized",
-                                        "Meta-Llama-3.1-8B")
     assert os.path.exists(
         llama_model_root
     ), f"{llama_model_root} does not exist under NFS LLM_MODELS_ROOT dir"

tests/integration/defs/disaggregated/test_ad_disagg.py

@@ -67,32 +67,13 @@ def skip_b300():
 AUTODEPLOY_DISAGG_SEED = 1234
 REDUCED_TINYLLAMA_LAYERS = 2
 REDUCED_DEEPSEEK_LAYERS = 2
-LLAMA_EAGLE3_EXPECTED_TEXT = " Berlin\nWhat is the capital of France? Paris\nWhat is the capital of"
-LLAMA_EAGLE3_EXPECTED_TOKEN_IDS = [
-    20437,
-    198,
-    3923,
-    374,
-    279,
-    6864,
-    315,
-    9822,
-    30,
-    12366,
-    198,
-    3923,
-    374,
-    279,
-    6864,
-    315,
-]
 
 
 MODEL_PATHS = {
-    "EAGLE3-LLaMA3.1-Instruct-8B": "EAGLE3-LLaMA3.1-Instruct-8B",
-    "Llama-3.1-8B-Instruct": "llama-3.1-model/Llama-3.1-8B-Instruct/",
     "TinyLlama-1.1B-Chat-v1.0": "llama-models-v2/TinyLlama-1.1B-Chat-v1.0",
     "DeepSeek-V3-Lite": "DeepSeek-V3-Lite/bf16",
+    "Qwen3-8B-eagle3": "Qwen3/qwen3_8b_eagle3",
+    "Qwen3-8B": "Qwen3/Qwen3-8B",
 }
 
 
@@ -657,14 +638,19 @@ def test_chunked_prefill_handoff(model):
 # ---------------------------------------------------------------------------
 
 
-def llama_eagle3_config():
+def qwen3_eagle3_config():
     return {
         "speculative_config": Eagle3DecodingConfig(
             max_draft_len=3,
-            speculative_model=model_path("EAGLE3-LLaMA3.1-Instruct-8B"),
-            eagle3_layers_to_capture={1, 15, 28},
+            speculative_model=model_path("Qwen3-8B-eagle3"),
+            # TODO: these capture layers were carried over proportionally from
+            # the retired Llama-3.1-8B config (1/32, 15/32, 28/32 through the
+            # stack) and have NOT been validated against Qwen3-8B's actual
+            # layer count on GPU. Re-derive and confirm before relying on this
+            # test's output.
+            eagle3_layers_to_capture={1, 17, 31},
         ),
-        # Force the Eagle3 draft to match the BF16 Llama 3.1 target. Shared KV
+        # Force the Eagle3 draft to match the BF16 Qwen3-8B target. Shared KV
         # cache management requires matching target and draft KV dtypes.
         "speculative_model_kwargs": {"torch_dtype": "bfloat16"},
     }
@@ -1039,18 +1025,39 @@ def test_async_sharded_generation_handoff():
 @pytest.mark.skip_less_device(2)
 @pytest.mark.timeout(900)
 def test_async_eagle3_full_model_handoff():
+    """Eagle3 one-model draft-token handoff, compared against an aggregate run.
+
+    Unlike the retired Llama-3.1-8B version of this test, this compares
+    against a freshly-computed aggregate (non-disaggregated) generation using
+    the same speculative_config, instead of hardcoded golden text/token IDs.
+    That avoids needing pre-recorded goldens for the new model pairing, at the
+    cost of also exercising the aggregate Eagle3 one-model path as a
+    dependency. This still needs a real GPU run to confirm Qwen3-8B +
+    Qwen3/qwen3_8b_eagle3 actually produce matching, non-trivial draft-token
+    output under this config (see the eagle3_layers_to_capture TODO in
+    qwen3_eagle3_config).
+    """
+    prompt = "What is the capital of Germany?"
     sampling_params_kwargs = {
         "max_tokens": 16,
         "ignore_eos": True,
         "top_k": 1,
         "seed": AUTODEPLOY_DISAGG_SEED,
     }
-    extra_config = llama_eagle3_config()
+    extra_config = qwen3_eagle3_config()
+
+    aggregate_output = run_aggregate_generation(
+        "Qwen3-8B",
+        world_size=1,
+        prompt=prompt,
+        sampling_params_kwargs=sampling_params_kwargs,
+        extra_config=extra_config,
+    )
     outputs = run_context_then_generation_handoff(
-        "Llama-3.1-8B-Instruct",
+        "Qwen3-8B",
         worker_world_sizes=(1, 1),
         generation_overlap=True,
-        prompt="What is the capital of Germany?",
+        prompt=prompt,
         sampling_params_kwargs=sampling_params_kwargs,
         extra_config=extra_config,
     )
@@ -1064,7 +1071,6 @@ def test_async_eagle3_full_model_handoff():
     assert outputs["generation"].token_ids
     assert has_draft_tokens(outputs["context"])
     assert has_draft_tokens(outputs["generation"])
-    assert outputs["context"].text == " Berlin"
-    assert outputs["context"].token_ids == LLAMA_EAGLE3_EXPECTED_TOKEN_IDS[:1]
-    assert outputs["generation"].text == LLAMA_EAGLE3_EXPECTED_TEXT
-    assert outputs["generation"].token_ids == LLAMA_EAGLE3_EXPECTED_TOKEN_IDS
+    assert outputs["context"].token_ids == aggregate_output.token_ids[:1]
+    assert outputs["generation"].text == aggregate_output.text
+    assert outputs["generation"].token_ids == aggregate_output.token_ids

tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_llama31_8b.yaml → tests/integration/defs/disaggregated/test_configs/disagg_config_ctxtp2_gentp2_qwen3_8b.yaml

@@ -1,5 +1,5 @@
 hostname: localhost
-model: llama-3.1-model/Llama-3.1-8B-Instruct
+model: Qwen3/Qwen3-8B
 free_gpu_memory_fraction: 0.25
 backend: pytorch
 disable_overlap_scheduler: true

tests/integration/defs/disaggregated/test_disaggregated.py

@@ -418,8 +418,8 @@ def get_test_config(test_desc, example_dir, test_root):
         f"{test_configs_root}/disagg_config_ctxtp2_gentp2_gptoss_tllm.yaml",
         "cancel_stress_test":
         f"{test_configs_root}/disagg_config_cancel_stress_test.yaml",
-        "llama31_8b":
-        f"{test_configs_root}/disagg_config_ctxtp2_gentp2_llama31_8b.yaml",
+        "qwen3_8b":
+        f"{test_configs_root}/disagg_config_ctxtp2_gentp2_qwen3_8b.yaml",
         "mamba_conc_greater_than_mbs":
         f"{test_configs_root}/disagg_config_mamba_conc_greater_than_mbs.yaml",
         "mamba_bs1_concurrency2":
@@ -2350,9 +2350,6 @@ def benchmark_model_root(request):
         model_path = os.path.join(models_root, "DeepSeek-V3-Lite", "fp8")
     elif (request.param == "DeepSeek-V3-Lite-bf16"):
         model_path = os.path.join(models_root, "DeepSeek-V3-Lite", "bf16")
-    elif request.param == "llama-3.1-8b-instruct-hf-fp8":
-        model_path = os.path.join(models_root, "llama-3.1-model",
-                                  "Llama-3.1-8B-Instruct-FP8")
     else:
         raise ValueError(f"Failed to find the model: {request.param}")
     return model_path
@@ -4074,11 +4071,8 @@ def test_disaggregated_cancel_large_context_requests(disaggregated_test_root,
 
 
 @pytest.mark.skip_less_device(4)
-@pytest.mark.parametrize("llama_model_root", ['llama-3.1-8b-instruct'],
-                         indirect=True)
 def test_disaggregated_logprobs_serving(disaggregated_test_root,
-                                        disaggregated_example_root, llm_venv,
-                                        llama_model_root):
+                                        disaggregated_example_root, llm_venv):
     """Test logprobs via OpenAI API in disaggregated serving with multi-GPU TP.
 
     Covers the RCCA scenario (NVBug 5926823): disaggregated + streaming + logprobs,
@@ -4140,24 +4134,25 @@ def extract_logprobs(result, api_type):
         logprobs = [item.get("logprob") for item in content]
         return tokens, logprobs
 
-    setup_model_symlink(llm_venv, llama_model_root,
-                        "llama-3.1-model/Llama-3.1-8B-Instruct")
+    model_path = "Qwen3/Qwen3-8B"
+    model_dir = f"{llm_models_root()}/{model_path}"
+    setup_model_symlink(llm_venv, model_dir, model_path)
 
-    config_file = get_test_config("llama31_8b", disaggregated_example_root,
+    config_file = get_test_config("qwen3_8b", disaggregated_example_root,
                                   os.path.dirname(__file__))
 
     env = llm_venv._new_env.copy()
     env["UCX_TLS"] = get_ucx_tls()
     ctx_workers, gen_workers, disagg_server, work_dir = [], [], None, None
     config, ctx_workers, gen_workers, disagg_server, server_port, work_dir = \
         setup_disagg_cluster(config_file, env=env,
-                             model_name=llama_model_root,
+                             model_name=model_dir,
                              cwd=llm_venv.get_working_directory(),
                              server_start_timeout=600)
 
     server_host = config.get("hostname", "localhost")
     server_url = f"http://{server_host}:{server_port}"
-    model_name = "llama-3.1-model/Llama-3.1-8B-Instruct"
+    model_name = model_path
     max_tokens = 20
     timeout = aiohttp.ClientTimeout(total=120)
     # Use emoji prompt to also stress-test multi-byte tokenizer handling

tests/integration/defs/disaggregated/test_disaggregated_single_gpu.py

@@ -56,9 +56,9 @@ def get_ucx_tls():
 MODEL_PATHS = {
     "DeepSeek-V3-Lite-fp8": "DeepSeek-V3-Lite/fp8",
     "TinyLlama-1.1B-Chat-v1.0": "llama-models-v2/TinyLlama-1.1B-Chat-v1.0",
-    "Llama-3.1-8B-Instruct": "llama-3.1-model/Llama-3.1-8B-Instruct/",
-    "EAGLE3-LLaMA3.1-Instruct-8B": "EAGLE3-LLaMA3.1-Instruct-8B",
+    "Qwen3-8B-eagle3": "Qwen3/qwen3_8b_eagle3",
     "Qwen3-8B-FP8": "Qwen3/Qwen3-8B-FP8",
+    "Qwen3-8B": "Qwen3/Qwen3-8B",
 }
 
 
@@ -517,8 +517,9 @@ def test_disaggregated_llama_context_capacity(model, enable_cuda_graph,
             print("All workers terminated.")
 
 
-@pytest.mark.parametrize("model", ["Llama-3.1-8B-Instruct"])
-@pytest.mark.parametrize("spec_dec_model_path", ["EAGLE3-LLaMA3.1-Instruct-8B"])
+@skip_pre_hopper
+@pytest.mark.parametrize("model", ["Qwen3-8B"])
+@pytest.mark.parametrize("spec_dec_model_path", ["Qwen3-8B-eagle3"])
 @pytest.mark.parametrize("generation_overlap", [False])
 def test_disaggregated_spec_dec_batch_slot_limit(model, spec_dec_model_path,
                                                  generation_overlap):