Revert unhelpful TF32 latency candidate

README.md

@@ -50,7 +50,6 @@ Mamba2 所需的 Triton runtime 以
 |---|---:|
 | CFG | fixed per-chunk schedule: row ordinal 0/even = 3.0, positive odd = 2.0; UI primary CFG is locked |
 | NFE steps | fixed at the validated value 10 |
-| CUDA matmul | deterministic TF32 on the pinned A10G runtime; cuBLAS workspace, cuDNN deterministic mode and runtime versions remain frozen |
 | Target pace | 4.0 speech units/sec |
 | Initial whole-trajectory policy | base: 5.2 CJK / 4.6 ASCII-mixed units/sec + 1 latent step |
 | Low-coverage chunk-refill policy | safe duration: 4.6 CJK / 4.0 ASCII-mixed units/sec + 1 latent step |

app.py

@@ -21,11 +21,8 @@ CUBLAS_WORKSPACE_CONFIG = ":4096:8"
 MAMBA_DETERMINISTIC = "1"
 CUDNN_BENCHMARK = False
 CUDNN_DETERMINISTIC = True
-# The pinned Ampere ZeroGPU runtime uses deterministic TF32 kernels.  This
-# preserves the deterministic release path while avoiding emulated full-FP32
-# tensor-core work on the hosted A10G.
-CUDA_MATMUL_ALLOW_TF32 = True
-CUDNN_ALLOW_TF32 = True
+CUDA_MATMUL_ALLOW_TF32 = False
+CUDNN_ALLOW_TF32 = False
 EXPECTED_TORCH_VERSION = "2.11.0+cu130"
 EXPECTED_TORCH_CUDA_VERSION = "13.0"
 EXPECTED_CUDNN_VERSION = 91900

tests/test_release_pins.py

@@ -221,8 +221,8 @@ def test_cuda_runtime_is_pinned_deterministic_before_model_import():
     assert constants["MAMBA_DETERMINISTIC"] == "1"
     assert constants["CUDNN_BENCHMARK"] is False
     assert constants["CUDNN_DETERMINISTIC"] is True
-    assert constants["CUDA_MATMUL_ALLOW_TF32"] is True
-    assert constants["CUDNN_ALLOW_TF32"] is True
+    assert constants["CUDA_MATMUL_ALLOW_TF32"] is False
+    assert constants["CUDNN_ALLOW_TF32"] is False
     assert constants["EXPECTED_TORCH_VERSION"] == "2.11.0+cu130"
     assert constants["EXPECTED_TORCH_CUDA_VERSION"] == "13.0"
     assert constants["EXPECTED_CUDNN_VERSION"] == 91900