[ci][diagnostic]: compare GameCraft T2V with FA4 off/on on GB200

.buildkite/scripts/lanes/ssim.sh

@@ -25,6 +25,20 @@ if [ "$selected" != all ]; then
   done
 fi
 
+# This disposable branch defaults to a focused GameCraft diagnostic. Preserve
+# the canonical path below so the lane contract remains intact and the branch
+# can be inspected locally with FASTVIDEO_GAMECRAFT_AB_DIAGNOSTIC=0.
+diagnostic_enabled=${FASTVIDEO_GAMECRAFT_AB_DIAGNOSTIC:-1}
+force_diagnostic_exit() {
+  trap - EXIT
+  exit 2
+}
+if [ "$diagnostic_enabled" = 1 ]; then
+  # Do not let a normal setup/test failure trigger the production SSIM retry
+  # policy; this experiment must produce one unambiguous pair of observations.
+  trap force_diagnostic_exit EXIT
+fi
+
 # MoGe's utils3d dependency builds glcontext from source on ARM64. The current
 # runner image predates the baked-in X11 headers below, so keep this guarded
 # bootstrap until every deployed image digest contains libx11-dev.
@@ -37,4 +51,91 @@ fi
 uv pip install git+https://github.com/microsoft/MoGe.git
 uv pip install k_diffusion einops_exts alias_free_torch torchsde
 
+if [ "$diagnostic_enabled" = 1 ]; then
+  probe_rcs=()
+  test_rcs=()
+  visible_gpus=${CUDA_VISIBLE_DEVICES:-0}
+  diagnostic_gpu=${visible_gpus%%,*}
+
+  run_variant() {
+    local fa4_requested=$1
+    local probe_rc
+    local test_rc
+    local -a variant_env=(
+      "CUDA_VISIBLE_DEVICES=$diagnostic_gpu"
+      "PYTORCH_CUDA_ALLOC_CONF=expandable_segments:False"
+      "FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN"
+      "FASTVIDEO_FA4=$fa4_requested"
+    )
+
+    echo "+++ GameCraft T2V diagnostic: FASTVIDEO_FA4=${fa4_requested}"
+    if env "${variant_env[@]}" python - <<'PY'
+import os
+
+import torch
+
+print(f"requested_backend={os.environ['FASTVIDEO_ATTENTION_BACKEND']}")
+print(f"requested_FASTVIDEO_FA4={os.environ['FASTVIDEO_FA4']}")
+print(f"torch_version={torch.__version__}")
+print(f"torch_cuda_version={torch.version.cuda}")
+print(f"cuda_available={torch.cuda.is_available()}")
+print(f"cuda_visible_devices={os.environ.get('CUDA_VISIBLE_DEVICES', '<unset>')}")
+print(f"gpu_count={torch.cuda.device_count()}")
+for index in range(torch.cuda.device_count()):
+    print(
+        f"gpu[{index}]={torch.cuda.get_device_name(index)} "
+        f"capability={torch.cuda.get_device_capability(index)}"
+    )
+
+try:
+    from fastvideo.attention.selector import get_attn_backend
+    from fastvideo.attention.utils.flash_attn_default import fa_version
+    from fastvideo.platforms import AttentionBackendEnum
+
+    requested_backend = AttentionBackendEnum[os.environ["FASTVIDEO_ATTENTION_BACKEND"]]
+    resolved_backend = get_attn_backend(
+        128,
+        torch.bfloat16,
+        supported_attention_backends=(AttentionBackendEnum.FLASH_ATTN,),
+        requested=requested_backend,
+    )
+except Exception as error:
+    print(f"resolved_backend=ERROR ({type(error).__name__}: {error})")
+    print("resolved_flash_attention=ERROR")
+    raise
+
+print(f"resolved_backend={resolved_backend.get_name()}")
+print(f"resolved_flash_attention=FA{fa_version}")
+PY
+    then
+      probe_rc=0
+    else
+      probe_rc=$?
+    fi
+
+    if env "${variant_env[@]}" python -m pytest \
+        fastvideo/tests/ssim/test_gamecraft_similarity.py::test_gamecraft_t2v_similarity \
+        -vs
+    then
+      test_rc=0
+    else
+      test_rc=$?
+    fi
+
+    probe_rcs+=("$probe_rc")
+    test_rcs+=("$test_rc")
+    echo "--- FASTVIDEO_FA4=${fa4_requested}: probe_rc=${probe_rc} test_rc=${test_rc}"
+  }
+
+  # Separate processes are required because FASTVIDEO_FA4 is resolved at
+  # import time. Run both variants even when either probe or test fails.
+  run_variant 0
+  run_variant 1
+
+  echo "+++ GameCraft T2V FA4-off/on diagnostic summary"
+  echo "FASTVIDEO_FA4=0 probe_rc=${probe_rcs[0]} test_rc=${test_rcs[0]}"
+  echo "FASTVIDEO_FA4=1 probe_rc=${probe_rcs[1]} test_rc=${test_rcs[1]}"
+  exit 2
+fi
+
 exec python fastvideo/tests/ssim/ci_runner.py "${args[@]}"