arron comments

Repositories
Issues
Comments

Results 4 comments of


                                            arron

Fully Async Recipe OOM with only 8192 response length

H100 80G ?

[recipe, tool] feat: support multi-turn and tool call for recipe/fully_async_policy

> @sl-1314 Hi, the wandb link https://wandb.ai/hou-zg-meituan/fully-async-policy-multiturn-tool?nw=nwuserhouzg there seems not public. Thanks for the reminder, you can check it again.

[recipe, tool] feat: support multi-turn and tool call for recipe/fully_async_policy

> @sl-1314 Hi, the wandb link https://wandb.ai/hou-zg-meituan/fully-async-policy-multiturn-tool?nw=nwuserhouzg there seems not public. Thanks for the reminder, you can check it again.

[WIP][sglang, recipe] feat: Add SGLang as rollout engine for one-step-off-policy with megatron

No obvious issues; it is similar to the implementation of fsdp_worker. Could you show the experimental results? @moehanabi