Fix CUDA int8 GEMM: 32-align pads, exact float fallback; remove triage code

app.py

@@ -206,12 +206,7 @@ def _encode_decode(audio_path: str, levels: tuple) -> dict:
         with torch.inference_mode():  # untimed warmup so the first row's RTF isn't inflated
             codec.sig_to_sig(sig)
             if NATIVE_LEVEL in focal_levels:
-                try:
-                    codec_i8.sig_to_sig(sig)
-                except NotImplementedError:  # TEMP triage
-                    import traceback
-
-                    raise gr.Error(traceback.format_exc()) from None
+                codec_i8.sig_to_sig(sig)
         for level in focal_levels:
             model = codec_i8 if level == NATIVE_LEVEL else codec
             if level != NATIVE_LEVEL:
@@ -219,12 +214,7 @@ def _encode_decode(audio_path: str, levels: tuple) -> dict:
             torch.cuda.synchronize()
             start = time.perf_counter()
             with torch.inference_mode():
-                try:
-                    rec = model.sig_to_sig(sig).squeeze(0).float().cpu().numpy()
-                except NotImplementedError:  # TEMP triage: surface the bare error's traceback
-                    import traceback
-
-                    raise gr.Error(traceback.format_exc()) from None
+                rec = model.sig_to_sig(sig).squeeze(0).float().cpu().numpy()
             torch.cuda.synchronize()
             rtf = (time.perf_counter() - start) / seconds
             rows.append(core.evaluate(ref, rec, sr, rtf, SIZES[level], level))
@@ -404,4 +394,4 @@ with gr.Blocks(title="FocalCodec quantization lab") as demo:
 
     gr.Markdown(FOOTER)
 
-demo.launch(mcp_server=True, theme=gr.themes.Soft(), show_error=True)  # show_error: temporary, remove after triage
+demo.launch(mcp_server=True, theme=gr.themes.Soft())

int8nn.py

@@ -17,7 +17,7 @@ import torch.nn.functional as F
 from torch import nn
 
 SCALE_PREFIX = "_scale/"
-_ROUND = 16  # torch._int_mm wants %16 dims on CUDA
+_ROUND = 32  # cublasLt int8 GEMM (via torch._int_mm) wants %32 dims on CUDA; %16 is not enough (K=16 fails)
 
 
 def _up(n: int) -> int:
@@ -37,14 +37,17 @@ def _act_quant(x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
 def _int_mm(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
     """(M, K) int8 x (K, N) int8 -> (M, N) int32. M is padded: CUDA rejects M <= 16."""
     m = a.shape[0]
-    mp = max(32, _up(m))
+    mp = max(_ROUND, _up(m))
     if mp != m:
         padded = torch.zeros((mp, a.shape[1]), dtype=a.dtype, device=a.device)
         padded[:m] = a
         a = padded
     try:
         return torch._int_mm(a, b)[:m]
-    except RuntimeError:  # backend refuses the shapes -> exact int32 matmul
+    except (RuntimeError, NotImplementedError):  # backend refuses the shape
+        if a.is_cuda:
+            # exact in float32 (|int8 * int8 * K| << 2^24); int32 matmul has no CUDA kernel
+            return (a.to(torch.float32) @ b.to(torch.float32))[:m]
         return (a.to(torch.int32) @ b.to(torch.int32))[:m]