feat: remove debug cuda avoid

server/text_generation_server/models/__init__.py

@@ -8,6 +8,7 @@
 from text_generation_server.utils.speculate import get_speculate, set_speculate
 from text_generation_server.models.model import Model
 from text_generation_server.models.causal_lm import CausalLM
+from text_generation_server.models.flash_causal_lm import FlashCausalLM
 from text_generation_server.models.bloom import BLOOMSharded
 from text_generation_server.models.mpt import MPTSharded
 from text_generation_server.models.seq2seq_lm import Seq2SeqLM
@@ -45,20 +46,6 @@
 
 FLASH_ATTENTION = True
 
-# FlashCausalLM reqiures CUDA Graphs to be enabled on the system. This will throw a RuntimeError
-# if CUDA Graphs are not available when calling `torch.cuda.graph_pool_handle()` in the FlashCausalLM
-HAS_CUDA_GRAPH = False
-try:
-    from text_generation_server.models.flash_causal_lm import FlashCausalLM
-
-    HAS_CUDA_GRAPH = True
-except RuntimeError as e:
-    logger.warning(f"Could not import FlashCausalLM: {e}")
-
-if HAS_CUDA_GRAPH:
-    __all__.append(FlashCausalLM)
-
-
 try:
     from text_generation_server.models.flash_rw import FlashRWSharded
     from text_generation_server.models.flash_neox import FlashNeoXSharded