fix(perf): default TF32 off; off = true fp32 (matmul + cuDNN conv)

Reported as "darker", but a fixed-seed spectral A/B shows TF32 is tonally
neutral (centroid 564→565 Hz, HF>8k 0.00825→0.00833) — the perceived change
is the seed=0 random-noise confound, not TF32. Still, TF32 is only ~1.15x and
not bit-exact, so default it OFF for reference-fp32 output and let compile
(~2.1x, op fusion) be the headline speedup. apply_tf32 now also toggles
cuDNN conv-TF32 (PyTorch leaves it on by default), so off is genuinely fp32.
Docs updated with the seed-confound A/B guidance.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

README.md

@@ -126,7 +126,7 @@ Loads (and caches) a checkpoint. Output: **`UNIVERSR_MODEL`**.
 |---|---|---|---|
 | `model` | choice | `universr-audio` | Preset to download, or a local checkpoint folder found under `models/universr/`. |
 | `device` | `auto` / `cuda` / `cpu` | `auto` | Where to load the weights. `auto` picks CUDA when available. |
-| `tf32` *(opt.)* | bool | `True` | TF32 matmul on Ampere+ (~1.15×). Perceptually lossless, not bit-exact. |
+| `tf32` *(opt.)* | bool | `False` | TF32 matmul + conv on Ampere+ (~1.15×). Tonally neutral in testing but not bit-exact; off = reference fp32. |
 | `compile` *(opt.)* | bool | `False` | `torch.compile` the network (~2×). See [Performance](#performance-speed). |
 | `local_path` *(opt.)* | string | `""` | Override: a folder with `config.yaml` + `pytorch_model.bin`, **or** a raw training checkpoint (`.pth` / `.ckpt`). |
 | `config_path` *(opt.)* | string | `""` | `config.yaml` to pair with a raw checkpoint. Empty → the bundled default config. |
@@ -228,16 +228,21 @@ Two ways to use it:
 
 ## Performance (speed)
 
-Two **equal-quality** speedups live on the Model Loader (both leave the output perceptually identical —
+Speedups live on the Model Loader. **`compile` is the real, tonally-neutral win** (op fusion); `tf32` is
-measured deviation is at the fp32 rounding floor, ≈ −64 dB):
+a small extra that is off by default.
 
 | Setting | Speedup (measured) | Notes |
 |---|---|---|
-| `tf32` (default **on**) | ~1.15× | TF32 matmul on Ampere+. One global flag, no caveats worth worrying about. |
+| `compile` (opt-in) | ~2.1× | `torch.compile` the network — op fusion, no tonal change. The recommended speedup. |
-| `compile` (opt-in) | ~2.1× | `torch.compile` the network. **Stacks with TF32 → ~2.5× total.** |
+| `tf32` (default **off**) | ~1.15× | TF32 matmul + conv on Ampere+. **Stacks with compile → ~2.5×.** Tonally neutral in our spectral A/B but not bit-exact — left off so the default is reference fp32. |
 
-On the reference machine, a 12 s clip went **4.3 s → 1.7 s (2.48×)** with both enabled, with a max
+On the reference machine, a 12 s clip went **4.3 s → 1.7 s (2.48×)** with both enabled.
-sample deviation of `2e-4` vs plain fp32.
+
+**About `tf32`:** in a fixed-seed A/B, TF32 left the spectral centroid and >8 kHz energy unchanged to 3
+significant figures (i.e. it does **not** darken the output). If you toggle it and the result sounds
+different, check your `seed` — with `seed=0` every run draws new noise, so two runs differ regardless of
+TF32. To compare fairly, set a fixed `seed` and change only the toggle. Enabling `tf32` also turns on
+cuDNN conv-TF32; disabling it restores true fp32 (PyTorch leaves conv-TF32 on by default otherwise).
 
 **About `compile`:** the first run pays a one-time compile (~10–35 s); after that the compiled model is
 cached for the whole ComfyUI session. The model can only be compiled for a **fixed input shape**, so the
@@ -245,10 +250,9 @@ node automatically **pads every chunk to `chunk_seconds`** — meaning clips of
 compiled graph (no per-length recompiles). Set the sampler's `chunk_seconds` near your typical clip length
 so short clips aren't padded up wastefully. Requires CUDA; falls back to eager if compilation fails.
 
-> These are the only speedups that don't change the output. Things that *don't* help here: CFG-batching,
+> Things that *don't* help here: CFG-batching, channel/chunk batching, and `channels_last` — the GPU is
-> channel/chunk batching, and `channels_last` — the GPU is already compute-bound at batch 1, so they
+> already compute-bound at batch 1, so they gave ~0 gain in testing. Going faster than `compile` requires
-> gave ~0 gain in testing. Going faster than this requires bf16/fp16, which is **not** equal-quality
+> bf16/fp16, which is **not** equal-quality (verify by ear first).
-> (verify by ear first).
 
 ## Recommended settings
 

nodes.py

@@ -56,9 +56,10 @@ class UniverSRModelLoader:
             },
             "optional": {
                 "tf32": ("BOOLEAN", {
-                    "default": True,
+                    "default": False,
-                    "tooltip": "Enable TF32 matmul on Ampere+ GPUs (~1.15x). Perceptually lossless "
+                    "tooltip": "TF32 matmul + conv on Ampere+ GPUs (~1.15x). Tonally neutral in testing "
-                               "but not bit-exact; global setting. Turn off for reference fp32.",
+                               "but not bit-exact; off by default = reference fp32. A/B with a FIXED seed "
+                               "(seed!=0) — comparing two seed=0 runs changes the noise, not just TF32.",
                 }),
                 "compile": ("BOOLEAN", {
                     "default": False,
@@ -83,7 +84,7 @@ class UniverSRModelLoader:
     RETURN_NAMES = ("model",)
     FUNCTION = "load"
 
-    def load(self, model, device, tf32=True, compile=False, local_path="", config_path=""):
+    def load(self, model, device, tf32=False, compile=False, local_path="", config_path=""):
         dev = _default_device() if device == "auto" else device
         if dev == "cuda" and not torch.cuda.is_available():
             print("[UniverSR] CUDA unavailable, falling back to CPU")

universr_wrapper.py

@@ -159,16 +159,21 @@ def resolve_model_ref(model: str, local_path: str = "") -> tuple:
 
 
 def apply_tf32(enabled: bool):
-    """Enable/disable TF32 matmul on Ampere+ GPUs. ~1.15x speedup, perceptually
+    """Enable/disable TF32 for BOTH matmul and cuDNN convolutions on Ampere+ GPUs.
-    lossless but NOT bit-exact (10 mantissa bits vs 23). Global process setting."""
+
+    ~1.15x when on. In our spectral A/B (centroid + HF energy) TF32 was tonally
+    neutral, but it is NOT bit-exact (10 mantissa bits vs 23), so it's off by
+    default. Off sets true fp32 — note PyTorch otherwise leaves cuDNN conv-TF32 ON
+    by default, so we explicitly disable it here too. Global process setting."""
     try:
-        torch.set_float32_matmul_precision("high" if enabled else "highest")
+        torch.set_float32_matmul_precision("high" if enabled else "highest")  # matmul TF32
+        torch.backends.cudnn.allow_tf32 = enabled                              # conv TF32
     except Exception:
         pass
 
 
 def load_model(model: str, device: str, local_path: str = "", config_path: str = "",
-               tf32: bool = True, compile_model: bool = False):
+               tf32: bool = False, compile_model: bool = False):
     """Load (and cache) a UniverSR model. Returns (model_obj, cache_key)."""
     apply_tf32(tf32)  # global; apply before the cache short-circuit so toggling takes effect
     kind, path = resolve_model_ref(model, local_path)