[None][ci] tighten VisualGen CBTS routing

Move remaining public VisualGen config imports to the public args module so _torch.visual_gen.config stays internal to VisualGen.

Treat the public tensorrt_llm/visual_gen package as outward-facing for CBTS fallback, while keeping internal _torch/visual_gen changes on VG-only narrowing with the current VG test-db families covered.

Signed-off-by: Zhenhua Wang <zhenhuaw@nvidia.com>

Author: zhenhuaw-me

jenkins/scripts/cbts/rules/README.md

@@ -163,32 +163,35 @@ Block selection — entry-pattern based only:
 VisualGen has no `condition.terms.backend` of its own; VG entries
 live in `backend: pytorch` and `backend: tensorrt` blocks. A block
 "belongs to VG" iff any of its `tests:` entries matches one of the
-three stable VG path families:
+stable VG path families:
 
-- `unittest/_torch/visual_gen/...` (28 entries)
-- `examples/test_visual_gen.py...` (1 entry)
-- `visual_gen/test_visual_gen_benchmark.py` (1 entry)
+- `visual_gen/...`
+- `examples/test_visual_gen.py...`
+- `examples/test_visual_gen_multi_gpu.py...`
+- `perf/test_visual_gen_perf_sanity.py...`
+- `unittest/media/test_encoding.py`
 
 For each matched block, `block_filters` keeps only the VG entries.
 Non-VG siblings in the same block stay governed by other rules.
 
-Outward-facing fallback: unlike AutoDeploy, VG is imported eagerly
-(top-level `from tensorrt_llm._torch.visual_gen.config import ...`
-in `commands/serve.py`, `commands/utils.py`,
-`serve/openai_server.py`). The 5 files that define / re-export the
-public API symbols (`VisualGenArgs`, `ParallelConfig`, `VisualGen`,
-`VisualGenParams`) are listed in `_VG_OUTWARD_FILES`; touching any
-of them claims the changed files but emits `scope=None` so Selector
-falls back to baseline. This protects trtllm-serve / trtllm-bench
-startup paths from VG signature drift slipping through pre-merge.
+Outward-facing fallback: unlike AutoDeploy, VG public symbols are
+imported eagerly by non-VG startup paths such as `commands/serve.py`,
+`commands/utils.py`, and `serve/openai_server.py`. The public API
+package prefix (`tensorrt_llm/visual_gen/`) is listed in
+`_VG_OUTWARD_PREFIXES`; touching any non-doc file under it claims the
+changed files but emits `scope=None` so Selector falls back to baseline.
+This protects trtllm-serve / trtllm-bench startup paths from VG
+signature drift slipping through pre-merge.
 
 Outcomes:
 
 - No VG source files in the diff → rule returns `None`.
-- VG source touched, all internal → `scope=visualgenonly`; sanity
-  off (VG changes don't affect wheel sanity); perfsanity on iff a
-  matched block lives in `l0_perf` or `*perf_sanity*`.
-- VG source touched, any outward-facing file → `scope=None`
+- VG source touched, all internal (`examples/visual_gen/**` or
+  `tensorrt_llm/_torch/visual_gen/**`) → `scope=visualgenonly`;
+  sanity off (VG changes don't affect wheel sanity); perfsanity on iff
+  a matched block lives in `l0_perf` or `*perf_sanity*`.
+- VG source touched, any outward-facing path under
+  `tensorrt_llm/visual_gen/` → `scope=None`
   (fallback).
 - VG source touched but no VG block found anywhere (defensive) →
   `scope=None` (fallback).

jenkins/scripts/cbts/rules/visual_gen_rule.py

@@ -21,20 +21,22 @@
 VisualGen does NOT have its own `condition.terms.backend`; VG test
 entries live in `backend: pytorch` and `backend: tensorrt` blocks.
 A block "belongs to VG" iff any of its `tests:` entries matches one
-of the three stable VG entry path families:
-  - `unittest/_torch/visual_gen/...`         (28 entries)
-  - `examples/test_visual_gen.py...`         (1 entry)
-  - `visual_gen/test_visual_gen_benchmark.py` (1 entry)
+of the stable VG entry path families:
+  - `visual_gen/...`
+  - `examples/test_visual_gen.py...`
+  - `examples/test_visual_gen_multi_gpu.py...`
+  - `perf/test_visual_gen_perf_sanity.py...`
+  - `unittest/media/test_encoding.py`
 
 Outward-facing fallback:
 Unlike AutoDeploy, VG is imported eagerly (module-level) by non-VG
 code: `commands/serve.py`, `commands/utils.py`, and
 `serve/openai_server.py` import `VisualGenArgs` / `ParallelConfig` /
 `VisualGen` / `VisualGenParams` at top level. A signature change to
 those symbols can break trtllm-serve startup, which would affect
-non-VG tests. The 5 files that define / re-export those symbols are
-listed in `_VG_OUTWARD_FILES`; touching any of them forces fallback
-even if the rest of the diff is VG-internal.
+non-VG tests. The public API package prefix is listed in
+`_VG_OUTWARD_PREFIXES`; touching any file under it forces fallback even
+if the rest of the diff is VG-internal.
 """
 
 from __future__ import annotations
@@ -55,27 +57,19 @@
     "tensorrt_llm/visual_gen/",
 )
 
-# Files inside _VG_SRC_PREFIXES that are imported eagerly by non-VG
-# code (top-level `from ... import VisualGenArgs / ParallelConfig /
-# VisualGen / VisualGenParams`). Touching any of these can break
-# trtllm-serve / trtllm-bench startup paths, so the rule defers to
-# baseline rather than narrowing.
-_VG_OUTWARD_FILES: frozenset[str] = frozenset(
-    {
-        "tensorrt_llm/_torch/visual_gen/config.py",
-        "tensorrt_llm/visual_gen/__init__.py",
-        "tensorrt_llm/visual_gen/args.py",
-        "tensorrt_llm/visual_gen/params.py",
-        "tensorrt_llm/visual_gen/visual_gen.py",
-    }
-)
+# Public VisualGen API package imported eagerly by non-VG code. Touching
+# any non-doc file under this prefix can break trtllm-serve / trtllm-bench
+# startup paths, so the rule defers to baseline rather than narrowing.
+_VG_OUTWARD_PREFIXES: tuple[str, ...] = ("tensorrt_llm/visual_gen/",)
 
-# Substrings that mark a test entry as VG. Cover all three path
-# families that appear in test-db YAMLs (audited 2026-05).
+# Substrings that mark a test entry as VG. Cover all VG path families
+# that appear in test-db YAMLs (audited 2026-06).
 _VG_ENTRY_PATTERNS: tuple[str, ...] = (
-    "unittest/_torch/visual_gen/",
+    "visual_gen/",
     "examples/test_visual_gen.py",
-    "visual_gen/test_visual_gen_benchmark.py",
+    "examples/test_visual_gen_multi_gpu.py",
+    "perf/test_visual_gen_perf_sanity.py",
+    "unittest/media/test_encoding.py",
 )
 
 
@@ -122,19 +116,19 @@ def apply(self, pr: PRInputs) -> Optional[RuleResult]:
         if not claimed:
             return None
 
-        # Outward-facing VG files break the "self-contained subsystem"
+        # Outward-facing VG paths break the "self-contained subsystem"
         # assumption — they are imported eagerly by trtllm-serve /
         # trtllm-bench. Claim the files (so they don't go unhandled and
         # silently fallback) but emit scope=None so Selector falls back
         # to baseline coverage instead of narrowing to VG-only stages.
-        outward = claimed & _VG_OUTWARD_FILES
+        outward = {f for f in claimed if f.startswith(_VG_OUTWARD_PREFIXES)}
         if outward:
             return RuleResult(
                 handled_files=claimed,
                 affected_stages=set(),
                 scope=None,
                 reason=(
-                    f"visualgen: {len(outward)} outward-facing VG file(s) "
+                    f"visualgen: {len(outward)} outward-facing VG path(s) "
                     f"touched ({sorted(outward)[0]}{'...' if len(outward) > 1 else ''}); "
                     "fallback to baseline"
                 ),

tensorrt_llm/bench/benchmark/visual_gen.py

@@ -199,8 +199,8 @@ def visual_gen_command(
     """Benchmark VisualGen (image/video generation) models offline."""
     import yaml
 
-    from tensorrt_llm._torch.visual_gen.config import VisualGenArgs
     from tensorrt_llm.visual_gen import VisualGen, VisualGenParams
+    from tensorrt_llm.visual_gen.args import VisualGenArgs
 
     if prompt is None and prompt_file is None:
         raise click.UsageError("Either --prompt or --prompt_file must be specified.")

tensorrt_llm/commands/utils.py

@@ -7,8 +7,8 @@
 import click
 from click.core import ParameterSource
 
-from tensorrt_llm._torch.visual_gen.config import ParallelConfig
 from tensorrt_llm.llmapi.utils import download_hf_partial
+from tensorrt_llm.visual_gen.args import ParallelConfig
 
 logger = logging.getLogger(__name__)
 

tests/integration/defs/examples/test_visual_gen_multi_gpu.py

@@ -40,8 +40,8 @@
 )
 
 try:
-    from tensorrt_llm._torch.visual_gen.config import ParallelConfig
     from tensorrt_llm._utils import get_free_port
+    from tensorrt_llm.visual_gen.args import ParallelConfig
 
     MODULES_AVAILABLE = True
 except ImportError:

tests/unittest/_torch/visual_gen/multi_gpu/test_tp_attention.py

@@ -40,11 +40,12 @@
 
 try:
     from tensorrt_llm._torch.device_mesh import DeviceMeshTopologyImpl
-    from tensorrt_llm._torch.visual_gen.config import AttentionConfig, DiffusionModelConfig
+    from tensorrt_llm._torch.visual_gen.config import DiffusionModelConfig
     from tensorrt_llm._torch.visual_gen.mapping import VisualGenMapping
     from tensorrt_llm._torch.visual_gen.modules.attention import Attention, QKVMode
     from tensorrt_llm._utils import get_free_port
     from tensorrt_llm.mapping import Mapping
+    from tensorrt_llm.visual_gen.args import AttentionConfig
 
     MODULES_AVAILABLE = True
 except ImportError:

tests/unittest/_torch/visual_gen/test_cosmos3_transformer.py

@@ -28,10 +28,10 @@
 import torch
 
 from tensorrt_llm._torch.modules.linear import Linear
-from tensorrt_llm._torch.visual_gen.config import DiffusionModelConfig, VisualGenArgs
+from tensorrt_llm._torch.visual_gen.config import DiffusionModelConfig
 from tensorrt_llm._torch.visual_gen.models.cosmos3.transformer_cosmos3 import Cosmos3VFMTransformer
 from tensorrt_llm._torch.visual_gen.pipeline_loader import PipelineComponent, PipelineLoader
-from tensorrt_llm.visual_gen.args import TorchCompileConfig
+from tensorrt_llm.visual_gen.args import TorchCompileConfig, VisualGenArgs
 
 pytestmark = pytest.mark.cosmos3
 

tests/unittest/_torch/visual_gen/test_qwen_image_registry.py

@@ -17,12 +17,13 @@
 # decorator on ``QwenImagePipeline`` being applied, which is what we are
 # testing here.
 from tensorrt_llm._torch.visual_gen import models  # noqa: F401
-from tensorrt_llm._torch.visual_gen.config import AttentionConfig, DiffusionModelConfig
+from tensorrt_llm._torch.visual_gen.config import DiffusionModelConfig
 from tensorrt_llm._torch.visual_gen.models.qwen_image import (
     QwenImagePipeline,
     QwenImageTransformer2DModel,
 )
 from tensorrt_llm._torch.visual_gen.pipeline_registry import PIPELINE_REGISTRY, AutoPipeline
+from tensorrt_llm.visual_gen.args import AttentionConfig
 
 
 def test_qwen_image_pipeline_is_registered():