[CI/Build][AMD] Add Llama4 Maverick FP8 to AMD CI (#28695)

Signed-off-by: zhewenli <zhewenli@meta.com>

.buildkite/lm-eval-harness/configs/Meta-Llama-4-Maverick-17B-128E-Instruct-FP8.yaml

@@ -8,3 +8,4 @@ tasks:
     value: 0.80
 limit: 250 # will run on 250 * 14 subjects = 3500 samples
 num_fewshot: 5
+rtol: 0.05

.buildkite/lm-eval-harness/configs/models-large-rocm.txt

@@ -0,0 +1 @@
+Meta-Llama-4-Maverick-17B-128E-Instruct-FP8.yaml

.buildkite/lm-eval-harness/test_lm_eval_correctness.py

@@ -9,11 +9,40 @@ pytest -s -v test_lm_eval_correctness.py \
     --tp-size=1
 """
 
+import os
+from contextlib import contextmanager
+
 import lm_eval
 import numpy as np
 import yaml
 
-RTOL = 0.08
+DEFAULT_RTOL = 0.08
+
+
+@contextmanager
+def scoped_env_vars(new_env: dict[str, str]):
+    if not new_env:
+        # Fast path: nothing to do
+        yield
+        return
+
+    old_values = {}
+    new_keys = []
+
+    try:
+        for key, value in new_env.items():
+            if key in os.environ:
+                old_values[key] = os.environ[key]
+            else:
+                new_keys.append(key)
+            os.environ[key] = str(value)
+        yield
+    finally:
+        # Restore / clean up
+        for key, value in old_values.items():
+            os.environ[key] = value
+        for key in new_keys:
+            os.environ.pop(key, None)
 
 
 def launch_lm_eval(eval_config, tp_size):
@@ -32,23 +61,26 @@ def launch_lm_eval(eval_config, tp_size):
         f"trust_remote_code={trust_remote_code},"
         f"max_model_len={max_model_len},"
     )
-    results = lm_eval.simple_evaluate(
-        model=backend,
-        model_args=model_args,
-        tasks=[task["name"] for task in eval_config["tasks"]],
-        num_fewshot=eval_config["num_fewshot"],
-        limit=eval_config["limit"],
-        # TODO(yeq): using chat template w/ fewshot_as_multiturn is supposed help
-        # text models. however, this is regressing measured strict-match for
-        # existing text models in CI, so only apply it for mm, or explicitly set
-        apply_chat_template=eval_config.get(
-            "apply_chat_template", backend == "vllm-vlm"
-        ),
-        fewshot_as_multiturn=eval_config.get("fewshot_as_multiturn", False),
-        # Forward decoding and early-stop controls (e.g., max_gen_toks, until=...)
-        gen_kwargs=eval_config.get("gen_kwargs"),
-        batch_size=batch_size,
-    )
+
+    env_vars = eval_config.get("env_vars", None)
+    with scoped_env_vars(env_vars):
+        results = lm_eval.simple_evaluate(
+            model=backend,
+            model_args=model_args,
+            tasks=[task["name"] for task in eval_config["tasks"]],
+            num_fewshot=eval_config["num_fewshot"],
+            limit=eval_config["limit"],
+            # TODO(yeq): using chat template w/ fewshot_as_multiturn is supposed help
+            # text models. however, this is regressing measured strict-match for
+            # existing text models in CI, so only apply it for mm, or explicitly set
+            apply_chat_template=eval_config.get(
+                "apply_chat_template", backend == "vllm-vlm"
+            ),
+            fewshot_as_multiturn=eval_config.get("fewshot_as_multiturn", False),
+            # Forward decoding and early-stop controls (e.g., max_gen_toks, until=...)
+            gen_kwargs=eval_config.get("gen_kwargs"),
+            batch_size=batch_size,
+        )
     return results
 
 
@@ -57,6 +89,8 @@ def test_lm_eval_correctness_param(config_filename, tp_size):
 
     results = launch_lm_eval(eval_config, tp_size)
 
+    rtol = eval_config.get("rtol", DEFAULT_RTOL)
+
     success = True
     for task in eval_config["tasks"]:
         for metric in task["metrics"]:
@@ -64,8 +98,9 @@ def test_lm_eval_correctness_param(config_filename, tp_size):
             measured_value = results["results"][task["name"]][metric["name"]]
             print(
                 f"{task['name']} | {metric['name']}: "
-                f"ground_truth={ground_truth} | measured={measured_value}"
+                f"ground_truth={ground_truth:.3f} | "
+                f"measured={measured_value:.3f} | rtol={rtol}"
             )
-            success = success and np.isclose(ground_truth, measured_value, rtol=RTOL)
+            success = success and np.isclose(ground_truth, measured_value, rtol=rtol)
 
     assert success

.buildkite/test-amd.yaml

@@ -718,17 +718,6 @@ steps:
   - uv pip install --system conch-triton-kernels
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 
-- label: LM Eval Small Models # 15min
-  timeout_in_minutes: 20
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_1
-  # grade: Blocking
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt --tp-size=1
-
 - label: OpenAI API correctness # 10min
   timeout_in_minutes: 15
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -974,19 +963,6 @@ steps:
     - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing
     - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
 
-- label: Multi-Modal Accuracy Eval (Small Models) # 10min
-  timeout_in_minutes: 70
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_1
-  # grade: Blocking
-  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
-  source_file_dependencies:
-  - vllm/multimodal/
-  - vllm/inputs/
-  - vllm/v1/core/
-  commands:
-  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1
-
 - label: Multi-Modal Models Test (Extended) 1 # 60min
   timeout_in_minutes: 120
   mirror_hardwares: [amdexperimental]
@@ -1162,21 +1138,6 @@ steps:
     # Run all e2e fusion tests
     - pytest -v -s tests/compile/distributed/test_fusions_e2e.py
 
-- label: ROCm GPT-OSS Eval
-  timeout_in_minutes: 60
-  working_dir: "/vllm-workspace/"
-  agent_pool: mi325_1
-  mirror_hardwares: [amdexperimental, amdproduction]
-  optional: true # run on nightlies
-  source_file_dependencies:
-  - tests/evals/gpt_oss
-  - vllm/model_executor/models/gpt_oss.py
-  - vllm/model_executor/layers/quantization/mxfp4.py
-  - vllm/v1/attention/backends/flashinfer.py
-  commands:
-    - uv pip install --system 'gpt-oss[eval]==0.0.5'
-    - VLLM_ROCM_USE_AITER_MHA=0 VLLM_ROCM_USE_AITER=1 VLLM_USE_AITER_UNIFIED_ATTENTION=1 pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
-
 - label: Blackwell Quantized MoE Test
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/"
@@ -1194,16 +1155,6 @@ steps:
   commands:
     - pytest -s -v tests/quantization/test_blackwell_moe.py
 
-- label: Blackwell LM Eval Small Models
-  timeout_in_minutes: 120
-  gpu: b200
-  optional: true # run on nightlies
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt --tp-size=1
-
 #####  1 GPU test  #####
 #####  multi gpus test  #####
 
@@ -1380,7 +1331,7 @@ steps:
     - pytest -v -s -x lora/test_llm_with_multi_loras.py
     - pytest -v -s -x lora/test_olmoe_tp.py
 
-    # Disabled for now because MXFP4 backend on non-cuda platform 
+    # Disabled for now because MXFP4 backend on non-cuda platform
     # doesn't support LoRA yet
     #- pytest -v -s -x lora/test_gptoss_tp.py
 
@@ -1446,37 +1397,6 @@ steps:
   - TARGET_TEST_SUITE=A100 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
   - pytest -v -s -x lora/test_mixtral.py
 
-- label: LM Eval Large Models # optional
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_4
-  # grade: Blocking
-  gpu: a100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
-
-##### H100 test #####
-- label: LM Eval Large Models (H100) # optional
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_4
-  # grade: Blocking
-  gpu: h100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  commands:
-    - export VLLM_USE_DEEP_GEMM=0  # We found Triton is faster than DeepGEMM for H100
-    - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-hopper.txt --tp-size=4
-
 ##### H200 test #####
 - label: Distributed Tests (H200) # optional
   mirror_hardwares: [amdexperimental]
@@ -1508,20 +1428,94 @@ steps:
     - pytest -v -s tests/distributed/test_nccl_symm_mem_allreduce.py
     - pytest -v -s tests/v1/distributed/test_dbo.py
 
-##### RL Integration Tests #####
-- label: Prime-RL Integration Test # 15min
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_2
+##### E2E Eval Tests #####
+- label: LM Eval Small Models (1 Card) # 15min
+  timeout_in_minutes: 20
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_1
   # grade: Blocking
-  timeout_in_minutes: 30
-  optional: true
-  num_gpus: 2
-  working_dir: "/vllm-workspace"
   source_file_dependencies:
-  - vllm/
-  - .buildkite/scripts/run-prime-rl-test.sh
+  - csrc/
+  - vllm/model_executor/layers/quantization
   commands:
-    - bash .buildkite/scripts/run-prime-rl-test.sh
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt --tp-size=1
+
+- label: Blackwell LM Eval Small Models
+  timeout_in_minutes: 120
+  gpu: b200
+  optional: true # run on nightlies
+  source_file_dependencies:
+  - csrc/
+  - vllm/model_executor/layers/quantization
+  commands:
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt --tp-size=1
+
+- label: Multi-Modal Accuracy Eval (Small Models) # 10min
+  timeout_in_minutes: 70
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_1
+  # grade: Blocking
+  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+  source_file_dependencies:
+  - vllm/multimodal/
+  - vllm/inputs/
+  - vllm/v1/core/
+  commands:
+  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1
+
+- label: LM Eval Large Models (4 Card)
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_4
+  # grade: Blocking
+  gpu: a100
+  optional: true
+  num_gpus: 4
+  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+  source_file_dependencies:
+  - csrc/
+  - vllm/model_executor/layers/quantization
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
+
+- label: LM Eval Large Models (H100) # optional
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_4
+  # grade: Blocking
+  gpu: h100
+  optional: true
+  num_gpus: 4
+  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+  source_file_dependencies:
+  - csrc/
+  - vllm/model_executor/layers/quantization
+  commands:
+    - export VLLM_USE_DEEP_GEMM=0  # We found Triton is faster than DeepGEMM for H100
+    - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-hopper.txt --tp-size=4
+
+- label: ROCm LM Eval Large Models (8 Card)
+  mirror_hardwares: [amdproduction]
+  agent_pool: mi325_8
+  num_gpus: 8
+  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-rocm.txt --tp-size=8
+
+- label: ROCm GPT-OSS Eval
+  timeout_in_minutes: 60
+  working_dir: "/vllm-workspace/"
+  agent_pool: mi325_1
+  mirror_hardwares: [amdexperimental, amdproduction]
+  optional: true # run on nightlies
+  source_file_dependencies:
+  - tests/evals/gpt_oss
+  - vllm/model_executor/models/gpt_oss.py
+  - vllm/model_executor/layers/quantization/mxfp4.py
+  - vllm/v1/attention/backends/flashinfer.py
+  commands:
+    - uv pip install --system 'gpt-oss[eval]==0.0.5'
+    - VLLM_ROCM_USE_AITER_MHA=0 VLLM_ROCM_USE_AITER=1 VLLM_USE_AITER_UNIFIED_ATTENTION=1 pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
 
 - label: DeepSeek V2-Lite Accuracy
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -1554,4 +1548,19 @@ steps:
   num_gpus: 2
   working_dir: "/vllm-workspace"
   commands:
-  - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
+  - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
+
+##### RL Integration Tests #####
+- label: Prime-RL Integration Test # 15min
+  mirror_hardwares: [amdexperimental]
+  agent_pool: mi325_2
+  # grade: Blocking
+  timeout_in_minutes: 30
+  optional: true
+  num_gpus: 2
+  working_dir: "/vllm-workspace"
+  source_file_dependencies:
+  - vllm/
+  - .buildkite/scripts/run-prime-rl-test.sh
+  commands:
+    - bash .buildkite/scripts/run-prime-rl-test.sh