From aaaeda98dcc14af75cd5fbd26386751a0e8b79c9 Mon Sep 17 00:00:00 2001
From: Divakar Verma <137818590+divakar-amd@users.noreply.github.com>
Date: Sat, 25 Jul 2026 00:15:19 -0400
Subject: [PATCH] [CI] fix compile test | refactor VLLM_DISABLE_COMPILE_CACHE
 for tests (#49770)

Signed-off-by: Divakar Verma <divakar.verma@amd.com>
---
 tests/compile/passes/test_fusion_attn.py         |  5 +----
 .../compile/passes/test_mla_attn_quant_fusion.py |  5 +----
 tests/compile/test_compile_ranges.py             | 11 +++++------
 tests/conftest.py                                | 16 ++++++++++++++++
 4 files changed, 23 insertions(+), 14 deletions(-)

diff --git a/tests/compile/passes/test_fusion_attn.py b/tests/compile/passes/test_fusion_attn.py
index 76536f3387c..da335b91557 100644
--- a/tests/compile/passes/test_fusion_attn.py
+++ b/tests/compile/passes/test_fusion_attn.py
@@ -290,12 +290,9 @@ def test_attention_quant_pattern(
     model_class: type[AttentionQuantPatternModel],
     backend: AttentionBackendEnum,
     dist_init,
-    monkeypatch,
-    use_fresh_inductor_cache,
+    disable_vllm_compile_cache,
 ):
     """Test AttentionStaticQuantPattern fusion pass"""
-    monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1")
-
     if backend == AttentionBackendEnum.FLASHINFER and (
         not current_platform.is_device_capability((10, 0)) or not has_flashinfer()
     ):
diff --git a/tests/compile/passes/test_mla_attn_quant_fusion.py b/tests/compile/passes/test_mla_attn_quant_fusion.py
index 0a38ffca483..5b2ef5bfd95 100644
--- a/tests/compile/passes/test_mla_attn_quant_fusion.py
+++ b/tests/compile/passes/test_mla_attn_quant_fusion.py
@@ -419,8 +419,7 @@ def test_mla_attention_quant_pattern(
     model_class: type[MLAAttentionQuantPatternModel],
     backend: AttentionBackendEnum,
     dist_init,
-    monkeypatch,
-    use_fresh_inductor_cache,
+    disable_vllm_compile_cache,
 ):
     """Test MLA AttentionQuantPattern fusion pass"""
     if (
@@ -429,8 +428,6 @@ def test_mla_attention_quant_pattern(
     ):
         pytest.skip("NVFP4 is not supported on this GPU (requires SM 100+).")
 
-    monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1")
-
     custom_ops_list = custom_ops.split(",") if custom_ops else []
 
     device = torch.device(f"{DEVICE_TYPE}:0")
diff --git a/tests/compile/test_compile_ranges.py b/tests/compile/test_compile_ranges.py
index 9fd8e9577ba..4dfea42a6b4 100644
--- a/tests/compile/test_compile_ranges.py
+++ b/tests/compile/test_compile_ranges.py
@@ -66,7 +66,7 @@ class PostGradRangeChecker(InductorPass):
         return InductorPass.hash_dict(state)
 
 
-def test_compile_ranges(use_fresh_inductor_cache):
+def test_compile_ranges(disable_vllm_compile_cache):
     post_grad_range_checker = PostGradRangeChecker(
         [
             Range(start=1, end=8),
@@ -168,7 +168,7 @@ class PostGradStaticShapeChecker(InductorPass):
         return InductorPass.hash_dict(state)
 
 
-def test_compile_sizes_produce_static_shapes(use_fresh_inductor_cache):
+def test_compile_sizes_produce_static_shapes(disable_vllm_compile_cache):
     """Verify that compile_sizes entries are compiled with fully concrete
     shapes (no SymInts), while compile_ranges entries retain dynamic shapes."""
     checker = PostGradStaticShapeChecker()
@@ -209,10 +209,9 @@ def test_compile_sizes_produce_static_shapes(use_fresh_inductor_cache):
     )
 
 
-def test_inductor_cache_compile_ranges(monkeypatch, use_fresh_inductor_cache):
-    # To force multiple compilations, we disable the compile cache
-    monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1")
-
+def test_inductor_cache_compile_ranges(disable_vllm_compile_cache):
+    # disable_vllm_compile_cache sets VLLM_DISABLE_COMPILE_CACHE=1 to force
+    # multiple compilations by disabling vLLM's on-disk compile cache.
     post_grad_range_checker = PostGradRangeChecker(
         ranges=[
             Range(start=1, end=8),
diff --git a/tests/conftest.py b/tests/conftest.py
index 47071167b56..406cb2ed2ea 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -1768,6 +1768,22 @@ def use_fresh_inductor_cache():
         yield
 
 
+@pytest.fixture
+def disable_vllm_compile_cache(monkeypatch, use_fresh_inductor_cache):
+    """
+    Use a fresh inductor cache AND disable vLLM's on-disk torch.compile cache.
+
+    This forces compilation (and any custom compile passes) to actually run
+    instead of being served from a warm cache left behind by previous runs
+    (e.g. on persistent CI agents). Use this for tests that inspect what
+    happens during compilation; use ``use_fresh_inductor_cache`` (or
+    ``fresh_vllm_cache``) instead when the vLLM compile cache must stay
+    enabled (e.g. cache save/load tests).
+    """
+    monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1")
+    yield
+
+
 @pytest.fixture
 def fresh_vllm_cache(monkeypatch, use_fresh_inductor_cache):
     """Temporary VLLM_CACHE_ROOT combined with a fresh inductor cache."""