diff --git a/tensorrt_llm/_torch/visual_gen/attention_backend/flash_attn4.py b/tensorrt_llm/_torch/visual_gen/attention_backend/flash_attn4.py
index 532e86e91e08..312aa13796b4 100644
--- a/tensorrt_llm/_torch/visual_gen/attention_backend/flash_attn4.py
+++ b/tensorrt_llm/_torch/visual_gen/attention_backend/flash_attn4.py
@@ -26,9 +26,11 @@
 
 from ...attention_backend.interface import PredefinedAttentionMask
 from .interface import AttentionBackend, AttentionTensorLayout
+from .utils import _merge_flash_attn_namespace
 
 _flash_attn_fwd_import_error = None
 try:
+    _merge_flash_attn_namespace()
     from flash_attn.cute.interface import _flash_attn_fwd
 except (ImportError, OSError) as e:
     _flash_attn_fwd = None
diff --git a/tensorrt_llm/_torch/visual_gen/attention_backend/parallel.py b/tensorrt_llm/_torch/visual_gen/attention_backend/parallel.py
index b9d5bac5e423..51091d3931ef 100644
--- a/tensorrt_llm/_torch/visual_gen/attention_backend/parallel.py
+++ b/tensorrt_llm/_torch/visual_gen/attention_backend/parallel.py
@@ -33,9 +33,11 @@
 
 from ...attention_backend.interface import PredefinedAttentionMask
 from .interface import AttentionBackend, AttentionTensorLayout
+from .utils import _merge_flash_attn_namespace
 
 _flash_attn_combine_import_error = None
 try:
+    _merge_flash_attn_namespace()
     from flash_attn.cute.interface import flash_attn_combine as _flash_attn_combine
 except (ImportError, OSError) as e:
     _flash_attn_combine = None
diff --git a/tensorrt_llm/_torch/visual_gen/attention_backend/utils.py b/tensorrt_llm/_torch/visual_gen/attention_backend/utils.py
index a3ce3df89cd9..8ac73770c8ec 100644
--- a/tensorrt_llm/_torch/visual_gen/attention_backend/utils.py
+++ b/tensorrt_llm/_torch/visual_gen/attention_backend/utils.py
@@ -30,6 +30,21 @@
 from .interface import AttentionBackend
 
 
+def _merge_flash_attn_namespace() -> None:
+    """Merge legacy `flash-attn` (v2) and `flash-attn-4` (cute-only) wheels.
+
+    `flash-attn-4` ships only `flash_attn/cute/` and relies on namespace
+    extension into the co-installed legacy `flash-attn` regular package.
+    When both wheels coexist, Python pins `flash_attn.__path__` to the v2
+    location and never sees `cute/`; re-extend the path to merge them.
+    """
+    import pkgutil
+
+    import flash_attn
+
+    flash_attn.__path__ = pkgutil.extend_path(flash_attn.__path__, flash_attn.__name__)
+
+
 def get_visual_gen_attention_backend(
     backend_name: str,
 ) -> Type[AttentionBackend]:
diff --git a/tests/integration/defs/examples/visual_gen/test_visual_gen.py b/tests/integration/defs/examples/visual_gen/test_visual_gen.py
index 916518acb644..365b079fde77 100644
--- a/tests/integration/defs/examples/visual_gen/test_visual_gen.py
+++ b/tests/integration/defs/examples/visual_gen/test_visual_gen.py
@@ -235,9 +235,13 @@ def _visual_gen_deps(llm_venv):
     """Install av + diffusers + ffmpeg once per session (shared by all video-gen fixtures)."""
     llm_venv.run_cmd(["-m", "pip", "install", "av"])
     llm_venv.run_cmd(["-m", "pip", "install", "diffusers>=0.37.0"])
-    # Install ffmpeg system package required by save_video() for MP4 encoding
-    check_call(["apt-get", "update", "-y"], shell=False)
-    check_call(["apt-get", "install", "-y", "ffmpeg"], shell=False)
+    if shutil.which("ffmpeg"):
+        return
+    # apt-get needs root; add `sudo` when the test harness is running as an
+    # unprivileged user (e.g. LOCAL_USER=1 dev containers) but sudo is available.
+    prefix = [] if os.geteuid() == 0 else ["sudo", "-n"]
+    check_call([*prefix, "apt-get", "update", "-y"], shell=False)
+    check_call([*prefix, "apt-get", "install", "-y", "ffmpeg"], shell=False)
 
 
 @pytest.fixture(scope="session")
diff --git a/tests/integration/defs/examples/visual_gen/test_visual_gen_multi_gpu.py b/tests/integration/defs/examples/visual_gen/test_visual_gen_multi_gpu.py
index a60765fe600c..8973191ac66c 100644
--- a/tests/integration/defs/examples/visual_gen/test_visual_gen_multi_gpu.py
+++ b/tests/integration/defs/examples/visual_gen/test_visual_gen_multi_gpu.py
@@ -15,6 +15,7 @@
 """Multi-GPU integration tests for VisualGen LPIPS quality checks."""
 
 import os
+import sys
 from typing import Callable
 
 import pytest
@@ -37,6 +38,7 @@
     _run_lpips_eval,
     _run_wan_lpips_pipeline,
     _save_lpips_video_mp4,
+    _visual_gen_deps,
 )
 
 try:
@@ -108,6 +110,11 @@ def run_test_in_distributed(world_size: int, test_fn: Callable, use_cuda: bool =
         pytest.skip(f"Test requires {world_size} GPUs, only {torch.cuda.device_count()} available")
     backend = "nccl" if use_cuda else "gloo"
     port = get_free_port()
+    # mp.spawn starts a fresh interpreter that does not inherit the parent's
+    # runtime sys.path mutations; without this, the child cannot import
+    # tensorrt_llm.bindings.internal.process_group (needed by the tp_size>1
+    # C++ allreduce path).
+    os.environ["PYTHONPATH"] = os.pathsep.join(filter(None, sys.path))
     mp.spawn(
         _distributed_worker,
         args=(world_size, backend, test_fn, port, kwargs),
@@ -158,7 +165,7 @@ def _wan22_lpips_distributed_worker(rank: int, world_size: int, **kwargs) -> Non
         dist.barrier()
 
 
-def _run_wan22_t2v_lpips_case(tmp_path, variant_name, parallel):
+def _run_wan22_t2v_lpips_case(_visual_gen_deps, tmp_path, variant_name, parallel):
     _skip_if_insufficient_gpus_for_parallel(parallel)
     parallel_cfg = ParallelConfig(**parallel)
     generated_path = tmp_path / f"wan22_t2v_generated_{variant_name}.mp4"
@@ -200,8 +207,8 @@ def _run_wan22_t2v_lpips_case(tmp_path, variant_name, parallel):
     WAN22_LPIPS_MULTI_GPU_VARIANTS,
     ids=[name for name, _ in WAN22_LPIPS_MULTI_GPU_VARIANTS],
 )
-def test_wan22_t2v_lpips_against_golden_multi_gpu(tmp_path, variant_name, parallel):
-    _run_wan22_t2v_lpips_case(tmp_path, variant_name, parallel)
+def test_wan22_t2v_lpips_against_golden_multi_gpu(_visual_gen_deps, tmp_path, variant_name, parallel):
+    _run_wan22_t2v_lpips_case(_visual_gen_deps, tmp_path, variant_name, parallel)
 
 
 @pytest.mark.parametrize(
@@ -209,5 +216,5 @@ def test_wan22_t2v_lpips_against_golden_multi_gpu(tmp_path, variant_name, parall
     WAN22_LPIPS_TP_VARIANTS,
     ids=[name for name, _ in WAN22_LPIPS_TP_VARIANTS],
 )
-def test_wan22_t2v_lpips_against_golden_tp(tmp_path, variant_name, parallel):
-    _run_wan22_t2v_lpips_case(tmp_path, variant_name, parallel)
+def test_wan22_t2v_lpips_against_golden_tp(_visual_gen_deps, tmp_path, variant_name, parallel):
+    _run_wan22_t2v_lpips_case(_visual_gen_deps, tmp_path, variant_name, parallel)
diff --git a/tests/integration/test_lists/waives.txt b/tests/integration/test_lists/waives.txt
index 11beee20e467..e42853fc2b9c 100644
--- a/tests/integration/test_lists/waives.txt
+++ b/tests/integration/test_lists/waives.txt
@@ -280,9 +280,6 @@ examples/visual_gen/test_visual_gen_multi_gpu.py::test_wan22_t2v_lpips_against_g
 examples/visual_gen/test_visual_gen_multi_gpu.py::test_wan22_t2v_lpips_against_golden_multi_gpu[cfg2_ulysses2] SKIP (https://nvbugs/6272644)
 examples/visual_gen/test_visual_gen_multi_gpu.py::test_wan22_t2v_lpips_against_golden_multi_gpu[cfg2_ulysses2_attn2d_2x1] SKIP (https://nvbugs/6272644)
 examples/visual_gen/test_visual_gen_multi_gpu.py::test_wan22_t2v_lpips_against_golden_multi_gpu[ulysses4] SKIP (https://nvbugs/6272644)
-examples/visual_gen/test_visual_gen_multi_gpu.py::test_wan22_t2v_lpips_against_golden_tp[cfg2_tp2] SKIP (https://nvbugs/6329227)
-examples/visual_gen/test_visual_gen_multi_gpu.py::test_wan22_t2v_lpips_against_golden_tp[tp2] SKIP (https://nvbugs/6329227)
-examples/visual_gen/test_visual_gen_multi_gpu.py::test_wan22_t2v_lpips_against_golden_tp[tp2_ulysses2] SKIP (https://nvbugs/6329227)
 full:A100/disaggregated/test_workers.py::test_workers_conditional_disaggregation_deepseek_v3_lite_bf16[DeepSeek-V3-Lite-bf16] SKIP (https://nvbugs/6329052)
 full:A100X/llmapi/test_llm_examples.py::test_llmapi_speculative_decoding_mtp SKIP (https://nvbugs/6287561)
 full:A100X/unittest/llmapi/test_llm_pytorch.py -m "part0" SKIP (https://nvbugs/6416249)