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>

Dockerfile.rocm

@@ -1,4 +1,4 @@
-FROM rocm/dev-ubuntu-22.04:7.2
+FROM rocm/dev-ubuntu-22.04:6.1
 
 ENV DEBIAN_FRONTEND=noninteractive \
     PYTHONDONTWRITEBYTECODE=1 \
@@ -29,12 +29,6 @@ RUN pip3 install -r requirements-rocm.txt
 # Step 3: Install chatterbox with --no-deps so pip cannot replace ROCm torch.
 RUN pip3 install --no-deps chatterbox-tts
 
-# Stub out resemble-perth (audio watermarking, unnecessary for self-hosted use).
-# Placed after pip layers so changes to the stub don't bust the cache above.
-COPY perth_stub.py .
-RUN python3 -c "import site; print(site.getsitepackages()[0])" | \
-    xargs -I{} cp /app/perth_stub.py {}/perth.py
-
 # Application source
 COPY engine.py config.py wyoming_handler.py wyoming_voices.py main.py ./
 

docker-compose.yml

@@ -28,18 +28,10 @@ services:
       - hf_cache:/app/hf_cache
     environment:
       - HF_HUB_ENABLE_HF_TRANSFER=1
-      # Required for RX 6700 XT (gfx1031) - not natively supported in ROCm 7.2.
+      # Required for RX 6700 XT (gfx1031) - not natively supported in ROCm.
       - HSA_OVERRIDE_GFX_VERSION=10.3.0
       # Disable MIOpen's SQLite cache — avoids crashes writing benchmark results.
-      # PyTorch's in-memory benchmark cache still applies within a container run.
       - MIOPEN_DISABLE_CACHE=1
-      # Disable MLIR-based ImplicitGEMM solvers. These compile MLIR kernels on the
-      # fly and hit 'too many open files' during the exhaustive benchmark search.
-      - MIOPEN_DEBUG_CONV_IMPLICIT_GEMM=0
-      # Suppress MIOpen workspace=0 fallback warnings (errors still shown).
-      # Levels: 0=quiet 1=fatal 2=error 3=warning(default) 4=info 5=debug
-      - MIOPEN_LOG_LEVEL=2
-      # - HF_TOKEN=your_token_here
 
 volumes:
   hf_cache:

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
 

main.py

@@ -18,33 +18,16 @@ logging.basicConfig(
 logger = logging.getLogger(__name__)
 
 
-_WARMUP_TEXTS = [
-    # Short: covers brief HA notifications (lights on/off, etc.)
-    "Okay.",
-    # Medium: covers typical HA announcements
-    "The front door is open. Please close it.",
-    # Long: covers longer TTS requests and pre-compiles dynamic shape graph
-    (
-        "This is a warmup synthesis to pre-compile neural network kernels "
-        "for longer text lengths used in Home Assistant announcements and notifications."
-    ),
-]
-
-
 def _warmup(voices: dict) -> None:
+    """Run one synthesis to populate MIOpen's in-memory kernel cache."""
     from wyoming_voices import resolve_voice
     audio_prompt = resolve_voice(None, voices) if voices else None
-    logger.info(
+    logger.info("Running warmup synthesis...")
-        f"Running {len(_WARMUP_TEXTS)}-pass warmup to pre-compile torch kernels "
-        "for short, medium, and long text lengths..."
-    )
-    for i, text in enumerate(_WARMUP_TEXTS, 1):
     try:
-            engine.synthesize(text=text, audio_prompt_path=audio_prompt)
+        engine.synthesize(text="Warmup.", audio_prompt_path=audio_prompt)
-            logger.info(f"Warmup pass {i}/{len(_WARMUP_TEXTS)} complete")
-        except Exception:
-            logger.warning(f"Warmup pass {i} failed (non-fatal)", exc_info=True)
         logger.info("Warmup complete")
+    except Exception:
+        logger.warning("Warmup synthesis failed (non-fatal)", exc_info=True)
 
 
 async def main() -> None:
@@ -58,10 +41,6 @@ async def main() -> None:
     voices = load_voices()
     wyoming_info = create_wyoming_info(engine.get_sample_rate(), voices)
 
-    # Run a warmup synthesis before accepting connections so MIOpen benchmarks
-    # and caches the best convolution algorithms for all layer shapes. Without
-    # this, the first real HA request triggers benchmarking (hundreds of runs)
-    # and times out before any audio is returned.
     _warmup(voices)
 
     host = get_wyoming_host()

requirements-rocm-init.txt

@@ -1,3 +1,5 @@
---index-url https://download.pytorch.org/whl/rocm7.2
+--index-url https://download.pytorch.org/whl/rocm6.1
-torch==2.11.0
+torch==2.5.1
-torchaudio==2.11.0
+torchaudio==2.5.1
+torchvision==0.20.1
+pytorch_triton_rocm==3.1.0

requirements-rocm.txt

@@ -5,18 +5,18 @@ librosa==0.11.0
 pyloudnorm
 
 # ML dependencies (pinned to match chatterbox without overwriting ROCm torch)
-transformers==5.2.0
+transformers==4.46.3
 diffusers==0.29.0
-safetensors==0.5.3
+safetensors==0.4.0
 huggingface-hub
 omegaconf
 
 # Chatterbox dependencies (installed separately since chatterbox uses --no-deps)
-# Note: resemble-perth is stubbed out in perth_stub.py (watermarking unneeded for self-hosted use)
 conformer==0.3.2
-s3tokenizer
+s3tokenizer==0.3.0
 spacy-pkuseg
 pykakasi==2.3.0
+resemble-perth==1.0.1
 
 # Wyoming protocol
 wyoming>=1.5.4