AI & ML interests
Internal testing artifact mangement for trl library
Recent Activity
View all activity
sergiopaniego
posted an update about 13 hours ago
Upload NemotronHForCausalLM
#1 opened 4 days ago
by
albertvillanova
Upload NemotronHForCausalLM
#1 opened 4 days ago
by
albertvillanova
Upload NemotronHForCausalLM
#1 opened 4 days ago
by
albertvillanova
Upload NemotronHForCausalLM
#1 opened 4 days ago
by
albertvillanova
Upload Qwen2_5_VLForConditionalGeneration
#20 opened 5 days ago
by
albertvillanova
Upload Qwen2VLForConditionalGeneration
#7 opened 5 days ago
by
albertvillanova
Upload GptOssForCausalLM
#4 opened 5 days ago
by
albertvillanova
Upload Lfm2VlForConditionalGeneration
#1 opened 5 days ago
by
albertvillanova
Upload DeepseekV3ForCausalLM
#4 opened 8 days ago
by
albertvillanova
Upload DeepseekV3ForCausalLM
#5 opened 8 days ago
by
albertvillanova
Upload Lfm2ForCausalLM
#1 opened 8 days ago
by
albertvillanova
Upload Phi3ForCausalLM
#2 opened 11 days ago
by
albertvillanova
Upload Phi3ForCausalLM
#2 opened 11 days ago
by
albertvillanova
Upload OPTForCausalLM
#3 opened 11 days ago
by
albertvillanova
Upload Gemma2ForCausalLM
#3 opened 11 days ago
by
albertvillanova
Upload GemmaForCausalLM
#3 opened 11 days ago
by
albertvillanova
Upload FalconMambaForCausalLM
#3 opened 11 days ago
by
albertvillanova
Upload Qwen2ForCausalLM
#2 opened 11 days ago
by
albertvillanova
sergiopaniego
posted an update 11 days ago
Post
347
ICYMI, Async GRPO in TRL now supports LoRA and we wrote a looong blog testing it
> the adapter is a few megabytes, so the weight sync is a file instead of an NCCL transfer
> 3 HF Jobs: 1 trainer and 2 vLLM replicas
> the adapter travels through an HF Storage Bucket mounted in all 3 at the same path
> a proxy in front of the replicas routes each rollout to the one already holding its KV prefix
https://huggingface.co/blog/asyncgrpo-lora-hfjobs
> the adapter is a few megabytes, so the weight sync is a file instead of an NCCL transfer
> 3 HF Jobs: 1 trainer and 2 vLLM replicas
> the adapter travels through an HF Storage Bucket mounted in all 3 at the same path
> a proxy in front of the replicas routes each rollout to the one already holding its KV prefix
https://huggingface.co/blog/asyncgrpo-lora-hfjobs