[Bugfix] Disable tma_aligned_scales in test_fusions_e2e (#32916)

Signed-off-by: Xin Yang <xyangx@amazon.com>

vllm/envs.py

@@ -162,6 +162,7 @@ if TYPE_CHECKING:
     VLLM_USE_DEEP_GEMM: bool = True
     VLLM_MOE_USE_DEEP_GEMM: bool = True
     VLLM_USE_DEEP_GEMM_E8M0: bool = True
+    VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES: bool = True
     VLLM_DEEP_GEMM_WARMUP: Literal[
         "skip",
         "full",
@@ -1201,6 +1202,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_USE_DEEP_GEMM_E8M0": lambda: bool(
         int(os.getenv("VLLM_USE_DEEP_GEMM_E8M0", "1"))
     ),
+    # Whether to create TMA-aligned scale tensor when DeepGEMM is used.
+    "VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES": lambda: bool(
+        int(os.getenv("VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES", "1"))
+    ),
     # DeepGemm JITs the kernels on-demand. The warmup attempts to make DeepGemm
     # JIT all the required kernels before model execution so there is no
     # JIT'ing in the hot-path. However, this warmup increases the engine

vllm/model_executor/layers/quantization/utils/fp8_utils.py

@@ -379,7 +379,7 @@ class W8A8BlockFp8LinearOp:
                 False,
                 self.act_quant_group_shape,
                 column_major_scales=True,
-                tma_aligned_scales=True,
+                tma_aligned_scales=envs.VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES,
                 use_ue8m0=self.use_deep_gemm_e8m0,
             )
             if self.is_deep_gemm_supported