Switch to ROCm 6.1 + torch 2.5.1 to fix MIOpen workspace=0 slowness

ROCm 7.2 + PyTorch 2.11.0 has a bug where PyTorch passes workspace=0 to
MIOpen convolutions, forcing fallback to the slow GemmFwdRest solver.
This caused s3gen.inference to take 15-22s instead of <5s, making
synthesis 3-4x slower than real-time audio playback.

ROCm 6.1 allocates workspace correctly so MIOpen picks fast GEMM solvers
without needing torch.compile workarounds.

Changes:
- Base image: rocm/dev-ubuntu-22.04:7.2 → 6.1
- torch 2.11.0 → 2.5.1 (rocm6.1 wheel index)
- Add pytorch_triton_rocm==3.1.0
- transformers 5.2.0 → 4.46.3, safetensors 0.5.3 → 0.4.0
- s3tokenizer unpinned → 0.3.0
- resemble-perth==1.0.1 directly (v1.0.1 is pip-installable; drop stub)
- Drop Dockerfile perth_stub steps
- Drop torch.compile and timing patches from engine.py (not needed)
- Drop multi-pass warmup from main.py (torch JIT warmup not needed)
- Drop ROCm 7.2-specific env vars from docker-compose.yml

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

engine.py

@@ -1,7 +1,5 @@
 import logging
-import time
 import torch
-import torch._dynamo
 from typing import Optional, Tuple
 
 logger = logging.getLogger(__name__)
@@ -46,23 +44,6 @@ def load_model() -> bool:
             _is_turbo = False
 
         _sample_rate = 24000
-
-        if torch.cuda.is_available():
-            # torch.compile replaces MIOpen's convolution path with Triton-generated
-            # kernels, bypassing the workspace=0 fallback entirely. We compile s3gen
-            # (HiFiGAN vocoder + flow matching) and the voice encoder (ve) since both
-            # are convolutional and hit the workspace=0 bug.
-            # suppress_errors=True falls back to eager for any op compile can't handle.
-            torch._dynamo.config.suppress_errors = True
-            for attr, label in [("s3gen", "s3gen"), ("ve", "ve")]:
-                try:
-                    obj = getattr(chatterbox_model, attr)
-                    setattr(chatterbox_model, attr, torch.compile(obj, dynamic=True))
-                    logger.info(f"{label} compiled with torch.compile")
-                except Exception:
-                    logger.warning(f"torch.compile unavailable for {label}, running in eager mode", exc_info=True)
-
-        _patch_timing(chatterbox_model)
         logger.info("Model loaded successfully")
         return True
     except Exception:
@@ -70,36 +51,6 @@ def load_model() -> bool:
         return False
 
 
-def _patch_timing(model) -> None:
-    """Wrap key sub-model forward() calls with timing logs."""
-    def _wrap(obj, method_name, label):
-        original = getattr(obj, method_name)
-        def timed(*args, **kwargs):
-            t0 = time.monotonic()
-            result = original(*args, **kwargs)
-            if torch.cuda.is_available():
-                torch.cuda.synchronize()
-            logger.info(f"[timing] {label}: {time.monotonic() - t0:.3f}s")
-            return result
-        setattr(obj, method_name, timed)
-
-    try:
-        # S3 tokenizer — processes reference audio through a conformer
-        _wrap(model.s3tokenizer, "forward", "s3tokenizer (ref audio encoding)")
-    except AttributeError:
-        pass
-    try:
-        # Speaker/voice encoder — xvector embedding from reference audio
-        _wrap(model.ve, "forward", "ve (speaker embedding)")
-    except AttributeError:
-        pass
-    try:
-        # S3Gen decode: flow matching (token -> mel) + HiFiGAN (mel -> wav)
-        _wrap(model.s3gen, "inference", "s3gen.inference (flow+vocoder)")
-    except AttributeError:
-        pass
-
-
 def get_sample_rate() -> int:
     return _sample_rate