Mistral Large 4 is out, with day-0 support in TRL 🔥
Fine-tune the 1T-parameter MoE with SFTTrainer + LoRA on a single node of 8 B300s, directly on top of its FP8 weights.
Example 👉 github.com/huggingface/trl/tree/main/examples/sft…
14 likes · 585 views
Mistral Large 4 (Le Chonk 🐈) is now live on the Hub and you can post-train it with TRL today!
SFT + LoRA example ready, runs on a single 8xB300 node
TRL example: github.com/huggingface/trl/tree/main/examples/sft… t.co
17 likes · 3 reposts · 1,170 views
TRL has day-0 support to post-train Le Chonk on a single node 🤯
LoRA SFT on a 1T-parameter model, training adapters on top of FP8 weights:
⚡ 32.7k tokens per step in ~8 seconds
💾 142 GiB per GPU on 8× B300
🔀 Expert parallelism + FSDP
Recipe: github.com/huggingface/trl/blob/ea936f801deeea713…
Open weights = open training after all ;)
32 likes · 3 reposts · 1,425 views
Beam, the new open model from @reflection_ai (501B parameters, 23B active), fine-tunes with LoRA and TRL's SFTTrainer on a single node of 8 B300s 🤯.
The whole parallelism setup is one line 👇
DistributedConfig(fsdp_size=8, ep_size=8)
github.com/huggingface/trl/tree/main/examples/sft… t.co
21 likes · 1 repost · 623 views
TRL v1.15 is out. Our biggest optimization ever.
Fused LM head, on by default:
→ up to 82% less peak VRAM
→ 7x longer sequences
→ DPO 10k → 59k tokens, GRPO 29k → 115k
And more... github.com/huggingface/trl/releases/tag/v1.15.0 t.co
38 likes · 8 reposts · 6,527 views
Mentions
Mistral Large 4 is out, with day-0 support in TRL 🔥 Fine-tune the 1T-parameter MoE with SFTTrainer + LoRA on a single node of 8 B300s, directly on top of its FP8 weights. Example 👉 github.com/huggingface/trl/tree/main/examples/sft…
Mistral Large 4 (Le Chonk 🐈) is now live on the Hub and you can post-train it with TRL today! SFT + LoRA example ready, runs on a single 8xB300 node TRL example: github.com/huggingface/trl/tree/main/examples/sft… t.co
TRL has day-0 support to post-train Le Chonk on a single node 🤯 LoRA SFT on a 1T-parameter model, training adapters on top of FP8 weights: ⚡ 32.7k tokens per step in ~8 seconds 💾 142 GiB per GPU on 8× B300 🔀 Expert parallelism + FSDP Recipe: github.com/huggingface/trl/blob/ea936f801deeea713… Open weights = open training after all ;)
Beam, the new open model from @reflection_ai (501B parameters, 23B active), fine-tunes with LoRA and TRL's SFTTrainer on a single node of 8 B300s 🤯. The whole parallelism setup is one line 👇 DistributedConfig(fsdp_size=8, ep_size=8) github.com/huggingface/trl/tree/main/examples/sft… t.co
TRL v1.15 is out. Our biggest optimization ever. Fused LM head, on by default: → up to 82% less peak VRAM → 7x longer sequences → DPO 10k → 59k tokens, GRPO 29k → 115k And more... github.com/huggingface/trl/releases/tag/v1.15.0 t.co