mirror of
https://github.com/allenai/olmocr.git
synced 2025-11-17 02:50:00 +00:00
Adding some more options to play with
This commit is contained in:
parent
755c221024
commit
2fd4ae8489
@ -806,8 +806,14 @@ def main():
|
|||||||
"--vllm_mode",
|
"--vllm_mode",
|
||||||
type=str,
|
type=str,
|
||||||
default="colocate",
|
default="colocate",
|
||||||
choices=["colocate", "server"],
|
choices=["colocate", "server", "none"],
|
||||||
help="VLLM execution mode: colocate or server (default: colocate)"
|
help="VLLM execution mode: colocate, server, or none to disable vllm (default: colocate)"
|
||||||
|
)
|
||||||
|
parser.add_argument(
|
||||||
|
"--num_iterations",
|
||||||
|
type=int,
|
||||||
|
default=1,
|
||||||
|
help="Number of GRPO iterations (default: 1)"
|
||||||
)
|
)
|
||||||
|
|
||||||
args = parser.parse_args()
|
args = parser.parse_args()
|
||||||
@ -955,10 +961,11 @@ def main():
|
|||||||
beta=args.beta,
|
beta=args.beta,
|
||||||
importance_sampling_level=args.importance_sampling_level,
|
importance_sampling_level=args.importance_sampling_level,
|
||||||
reward_weights=reward_weights,
|
reward_weights=reward_weights,
|
||||||
|
num_iterations=args.num_iterations,
|
||||||
|
|
||||||
# Vllm setup to speed up generation
|
# Vllm setup to speed up generation
|
||||||
use_vllm=True,
|
use_vllm=(args.vllm_mode != "none"),
|
||||||
vllm_mode=args.vllm_mode,
|
vllm_mode=args.vllm_mode if args.vllm_mode != "none" else "colocate",
|
||||||
vllm_gpu_memory_utilization=0.15,
|
vllm_gpu_memory_utilization=0.15,
|
||||||
log_completions=True,
|
log_completions=True,
|
||||||
)
|
)
|
||||||
|
|||||||
Loading…
x
Reference in New Issue
Block a user