feat: enable FlashAttention-2 for all models via runtime install + SDPA fallback

- app.py: runtime pip install flash-attn at startup (before model loading)
- voice_clone.py: pass attn_implementation to both TTS models, fallback to SDPA
- inference.py: enable FA2/SDPA for Qwen2.5-3B-Instruct
- inference_lfm.py: enable FA2/SDPA for LFM2.5-Audio
- requirements_lfm.txt: removed flash-attn (can't compile during Docker build)

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

app.py

@@ -1,3 +1,11 @@
+import subprocess, sys
+
+# Install FlashAttention-2 at runtime (needs CUDA, can't compile during Docker build)
+subprocess.run(
+    [sys.executable, "-m", "pip", "install", "flash-attn", "--no-build-isolation", "-q"],
+    capture_output=True,
+)
+
 import os
 import html
 import math

inference.py

@@ -78,6 +78,13 @@ def get_qa_model():
 
         if use_gpu:
             load_kwargs = {"device_map": "auto", "torch_dtype": torch.float16}
+            # Enable FlashAttention-2 if available, else SDPA
+            try:
+                import flash_attn  # noqa: F401
+                load_kwargs["attn_implementation"] = "flash_attention_2"
+                logger.info("Using FlashAttention-2 for Q&A model.")
+            except ImportError:
+                load_kwargs["attn_implementation"] = "sdpa"
             try:
                 from transformers import BitsAndBytesConfig
                 load_kwargs["quantization_config"] = BitsAndBytesConfig(

inference_lfm.py

@@ -25,7 +25,15 @@ def get_lfm_model():
 
         logger.info("Loading %s...", HF_REPO)
         _processor = LFM2AudioProcessor.from_pretrained(HF_REPO).eval()
-        _model = LFM2AudioModel.from_pretrained(HF_REPO).eval()
+        # Enable FlashAttention-2 if available, else SDPA
+        model_kwargs = {}
+        try:
+            import flash_attn  # noqa: F401
+            model_kwargs["attn_implementation"] = "flash_attention_2"
+            logger.info("Using FlashAttention-2 for LFM.")
+        except ImportError:
+            model_kwargs["attn_implementation"] = "sdpa"
+        _model = LFM2AudioModel.from_pretrained(HF_REPO, **model_kwargs).eval()
         logger.info("LFM2.5-Audio loaded.")
     return _processor, _model
 

requirements_lfm.txt

@@ -15,6 +15,6 @@ liquid-audio
 qwen-tts>=0.1.1
 huggingface-hub>=0.23.0
 
-# FlashAttention-2 — install separately on GPU machines:
-#   pip install flash-attn --no-build-isolation
-# Falls back to SDPA (PyTorch native) if not available.
+# FlashAttention-2: not in requirements (needs CUDA at build time).
+# Code auto-falls back to eager/SDPA if unavailable.
+# To enable: pip install flash-attn --no-build-isolation (at runtime)

voice_clone.py

@@ -99,13 +99,13 @@ def _select_dtype() -> torch.dtype:
 
 
 def _select_attn_impl() -> str:
-    """Use FlashAttention-2 if available, else eager (safest fallback)."""
+    """Use FlashAttention-2 if available, else SDPA (PyTorch native)."""
     try:
         import flash_attn  # noqa: F401
         return "flash_attention_2"
     except ImportError:
-        logger.info("flash-attn not installed — using eager attention.")
-        return "eager"
+        logger.info("flash-attn not installed — using SDPA attention.")
+        return "sdpa"
 
 
 def get_qwen_tts():
@@ -120,11 +120,13 @@ def get_qwen_tts():
                 device = "cuda" if torch.cuda.is_available() else "cpu"
 
                 logger.info("Loading %s on %s (cloning model)...", model_id, device)
+                attn_impl = _select_attn_impl()
                 _qwen_tts_model = Qwen3TTSModel.from_pretrained(
                     model_id,
                     device_map=device,
+                    attn_implementation=attn_impl,
                 )
-                logger.info("Qwen3-TTS Base loaded on %s.", device)
+                logger.info("Qwen3-TTS Base loaded on %s (attn=%s).", device, attn_impl)
     return _qwen_tts_model
 
 
@@ -139,11 +141,13 @@ def get_custom_voice_model():
                 device = "cuda" if torch.cuda.is_available() else "cpu"
 
                 logger.info("Loading CustomVoice 0.6B on %s...", device)
+                attn_impl = _select_attn_impl()
                 _custom_voice_model = Qwen3TTSModel.from_pretrained(
                     CUSTOM_VOICE_MODEL_ID,
                     device_map=device,
+                    attn_implementation=attn_impl,
                 )
-                logger.info("CustomVoice 0.6B loaded.")
+                logger.info("CustomVoice 0.6B loaded (attn=%s).", attn_impl)
     return _custom_voice_model