Commit
·
b4a7f49
1
Parent(s):
90312a8
Add Flash Attention 3 support with fallback to eager
Browse files- gpt_oss_transformers.py +25 -14
gpt_oss_transformers.py
CHANGED
@@ -264,13 +264,16 @@ def main(
|
|
264 |
logger.warning("Consider using --flavor a10g-large (48GB) or a100-large (80GB)")
|
265 |
|
266 |
try:
|
|
|
|
|
267 |
model = AutoModelForCausalLM.from_pretrained(
|
268 |
model_id,
|
269 |
torch_dtype=torch.bfloat16,
|
270 |
-
attn_implementation="
|
271 |
**model_kwargs,
|
272 |
)
|
273 |
model.eval()
|
|
|
274 |
except torch.cuda.OutOfMemoryError as e:
|
275 |
logger.error(f"Out of memory error: {e}")
|
276 |
logger.error("\nThe GPT OSS models require significant memory:")
|
@@ -278,19 +281,27 @@ def main(
|
|
278 |
logger.error("- 120B model: ~240GB VRAM (use 4xa100 or 8xl40s)")
|
279 |
sys.exit(1)
|
280 |
except Exception as e:
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
294 |
sys.exit(1)
|
295 |
|
296 |
# Generation configuration
|
|
|
264 |
logger.warning("Consider using --flavor a10g-large (48GB) or a100-large (80GB)")
|
265 |
|
266 |
try:
|
267 |
+
# Try Flash Attention 3 first for better performance
|
268 |
+
logger.info("Attempting to load with Flash Attention 3 (optimized)")
|
269 |
model = AutoModelForCausalLM.from_pretrained(
|
270 |
model_id,
|
271 |
torch_dtype=torch.bfloat16,
|
272 |
+
attn_implementation="kernels-community/vllm-flash-attn3", # Flash Attention with Sinks
|
273 |
**model_kwargs,
|
274 |
)
|
275 |
model.eval()
|
276 |
+
logger.info("Successfully loaded with Flash Attention 3")
|
277 |
except torch.cuda.OutOfMemoryError as e:
|
278 |
logger.error(f"Out of memory error: {e}")
|
279 |
logger.error("\nThe GPT OSS models require significant memory:")
|
|
|
281 |
logger.error("- 120B model: ~240GB VRAM (use 4xa100 or 8xl40s)")
|
282 |
sys.exit(1)
|
283 |
except Exception as e:
|
284 |
+
if "kernels-community" in str(e) or "flash" in str(e).lower() or "attn" in str(e).lower():
|
285 |
+
logger.warning(f"Flash Attention 3 not available: {type(e).__name__}")
|
286 |
+
logger.info("Falling back to eager attention (standard implementation)")
|
287 |
+
try:
|
288 |
+
model = AutoModelForCausalLM.from_pretrained(
|
289 |
+
model_id,
|
290 |
+
torch_dtype=torch.bfloat16,
|
291 |
+
attn_implementation="eager", # Fallback to eager attention
|
292 |
+
**model_kwargs,
|
293 |
+
)
|
294 |
+
model.eval()
|
295 |
+
logger.info("Successfully loaded with eager attention")
|
296 |
+
except torch.cuda.OutOfMemoryError as oom_error:
|
297 |
+
logger.error(f"Out of memory with eager attention: {oom_error}")
|
298 |
+
logger.error("The model requires more GPU memory than available")
|
299 |
+
sys.exit(1)
|
300 |
+
except Exception as eager_error:
|
301 |
+
logger.error(f"Failed with eager attention: {eager_error}")
|
302 |
+
sys.exit(1)
|
303 |
+
else:
|
304 |
+
logger.error(f"Unexpected error loading model: {e}")
|
305 |
sys.exit(1)
|
306 |
|
307 |
# Generation configuration
|