[None][feat] AutoDeploy: Fix hardcoded configs

tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py

@@ -393,11 +393,6 @@ def __init__(
         self.llm_args.print_iter_log = reporting_info.print_log
         self.llm_args.enable_iter_perf_stats = reporting_info.enable_iter_perf_stats
         self.llm_args.enable_iter_req_stats = reporting_info.enable_iter_req_stats
-        self.llm_args.stream_interval = 1
-        self.llm_args.attention_dp_config = None
-        self.llm_args.batch_wait_timeout_ms = 0
-        self.llm_args.batch_wait_timeout_iters = 0
-        self.llm_args.batch_wait_max_tokens_ratio = 0.0
         self.llm_args.max_num_tokens = cache_seq_interface.info.max_num_tokens
         self.llm_args.max_seq_len = cache_seq_interface.info.max_seq_len
         self.iter_counter = 0
@@ -407,12 +402,22 @@ def __init__(
         self.enable_attention_dp = dist_config.enable_attention_dp if dist_config else False
 
         if ad_config is not None:
+            self.llm_args.stream_interval = ad_config.stream_interval
+            self.llm_args.attention_dp_config = ad_config.attention_dp_config
+            self.llm_args.batch_wait_timeout_ms = ad_config.batch_wait_timeout_ms
+            self.llm_args.batch_wait_timeout_iters = ad_config.batch_wait_timeout_iters
+            self.llm_args.batch_wait_max_tokens_ratio = ad_config.batch_wait_max_tokens_ratio
             self.max_beam_width = ad_config.max_beam_width
             self.spec_config = ad_config.speculative_config
             self._disable_overlap_scheduler = ad_config.disable_overlap_scheduler
             self.llm_args.max_stats_len = ad_config.max_stats_len
             self._enable_chunked_prefill = getattr(ad_config, "enable_chunked_prefill", False)
         else:
+            self.llm_args.stream_interval = 1
+            self.llm_args.attention_dp_config = None
+            self.llm_args.batch_wait_timeout_ms = 0
+            self.llm_args.batch_wait_timeout_iters = 0
+            self.llm_args.batch_wait_max_tokens_ratio = 0.0
             self.max_beam_width = 1
             self.spec_config = None
             self._disable_overlap_scheduler = False

tests/unittest/auto_deploy/singlegpu/shim/test_engine.py

@@ -21,10 +21,12 @@
 
 from tensorrt_llm import SamplingParams
 from tensorrt_llm._torch.auto_deploy._compat import KvCacheConfig
+from tensorrt_llm._torch.auto_deploy.llm_args import LlmArgs
 from tensorrt_llm._torch.auto_deploy.shim.ad_executor import ADEngine
 from tensorrt_llm._torch.auto_deploy.shim.demollm import DemoEngine
 from tensorrt_llm._torch.auto_deploy.shim.interface import CachedSequenceInterface
 from tensorrt_llm._torch.pyexecutor.scheduler import ScheduledRequests
+from tensorrt_llm.llmapi import AttentionDpConfig
 
 
 class TransformerLikeModelwithFakeCachePool(nn.Module):
@@ -107,6 +109,82 @@ def test_engine(engine_cls: Type[ADEngine], tokens_per_block: int):
     cache_seq_interface.shutdown()
 
 
+def _make_cache_seq_interface(device, max_seq_len=64, max_batch_size=8):
+    return CachedSequenceInterface(
+        max_seq_len=max_seq_len,
+        max_batch_size=max_batch_size,
+        max_num_tokens=default_max_num_tokens(max_seq_len, max_batch_size),
+        device=device,
+        kv_cache_config=KvCacheConfig(tokens_per_block=max_seq_len),
+    )
+
+
+def test_ad_engine_propagates_pyexecutor_scheduling_config():
+    """ADEngine must copy PyExecutor scheduling/streaming knobs from ad_config.
+
+    Regression: these were hardcoded stubs (stream_interval=1,
+    attention_dp_config=None, batch_wait_*=0), silently dropping the yaml settings.
+    """
+    if not torch.cuda.is_available():
+        pytest.skip("ADEngine construction builds the model on CUDA")
+
+    device = torch.device("cuda")
+    max_seq_len, max_batch_size = 64, 8
+    cache_seq_interface = _make_cache_seq_interface(device, max_seq_len, max_batch_size)
+    cache_seq_interface.to(device)
+
+    ad_config = LlmArgs(
+        model="test-model",
+        max_batch_size=max_batch_size,
+        max_seq_len=max_seq_len,
+        max_input_len=32,
+        backend="_autodeploy",
+        cuda_graph_config={"max_batch_size": max_batch_size},
+        stream_interval=20,
+        attention_dp_config=AttentionDpConfig(
+            enable_balance=True, batching_wait_iters=50, timeout_iters=1
+        ),
+        batch_wait_timeout_ms=5.0,
+        batch_wait_timeout_iters=3,
+        batch_wait_max_tokens_ratio=0.5,
+    )
+
+    try:
+        engine = ADEngine(get_inference_model, cache_seq_interface, ad_config=ad_config)
+
+        assert engine.llm_args.stream_interval == 20
+        assert engine.llm_args.attention_dp_config is not None
+        assert engine.llm_args.attention_dp_config.enable_balance is True
+        assert engine.llm_args.attention_dp_config.batching_wait_iters == 50
+        assert engine.llm_args.attention_dp_config.timeout_iters == 1
+        assert engine.llm_args.batch_wait_timeout_ms == 5.0
+        assert engine.llm_args.batch_wait_timeout_iters == 3
+        assert engine.llm_args.batch_wait_max_tokens_ratio == 0.5
+    finally:
+        cache_seq_interface.shutdown()
+
+
+def test_ad_engine_scheduling_config_defaults_without_ad_config():
+    """Without ad_config, ADEngine falls back to the previous stub defaults."""
+    if not torch.cuda.is_available():
+        pytest.skip("ADEngine construction builds the model on CUDA")
+
+    device = torch.device("cuda")
+    cache_seq_interface = _make_cache_seq_interface(device)
+    cache_seq_interface.to(device)
+
+    try:
+        engine = ADEngine(get_inference_model, cache_seq_interface)
+
+        assert engine.llm_args.stream_interval == 1
+        assert engine.llm_args.attention_dp_config is None
+        assert engine.llm_args.batch_wait_timeout_ms == 0
+        assert engine.llm_args.batch_wait_timeout_iters == 0
+        assert engine.llm_args.batch_wait_max_tokens_ratio == 0.0
+    finally:
+        cache_seq_interface.shutdown()
+
+
 @pytest.mark.parametrize("tokens_per_block", [0, 2])
 def test_demo_engine_sampling(tokens_per_block: int):
     """Test sampling logic specific to DemoEngine."""