nullHawk

Here is something that ma boi has been cookin for a while... it's realy impressive!!!

@Gauri_the_great is this ASR or speech model?

Do I have a SOTA model ?

type shi one has to do for data

@WhispererCode 🤫

@null_hawk Whats your reward model ?

Its been approx 20+ hrs and finally reduced my GRPO runtime from ~12.8 hrs to ~1.5 hrs. used vLLM inference on one GPU and DDP on the other 3, applied pretty much everything I had. using 3 GPUs for DDP, hosted vLLM inference on one GPU, with 8 samples generating parallely, along with codec tokens, grad_accum=4, data sharding across ranks, micro-batching within the loss function, TF32 tensor cores, BF16 autocast for forward pass, Flash SDP + mem-efficient SDP, cuDNN benchmark mode, BF16 reduced precision reduction... works well for Hopper architecture. also turned on gradient checkpointing, manual gradient all-reduce (instead of full DDP wrapping), plus a file-based vLLM watcher that restarts the inference server with fresh merged weights at every optimizer step from a clean process tree (to avoid NCCL conflicts with torchrun), with retry logic on generation calls during server. biggest debugging rabbit holes: vLLM V1 silently ignoring stop_token_ids (had to force V0 with VLLM_USE_V1=0), and merge_and_unload() with tie_word_embeddings=True dropping the trained lm_head during save, model generates infinite codec tokens and never stops. fix: untie before merge so both embed_tokens and lm_head are saved separately. GRPO on TTS is a different beast from text!

though 50% of the time is spent by reward model

@cneuralnetwork @smallest_AI i don't know why it didn't worked when i tried it... maybe its upset with me :(

okay so lightning v3.1 from @smallest_AI gets hard hindi tongue twisters right

@atulit_gaur Nah bro... I don't believe my college

@null_hawk I was supposed to apply 🤣

anybody else's personal dms look like this? i have obviously not opened these links

just call the "LLM" a "Policy" voila!!! you are an RL guy now

@SherryYanJiang @amilabs @theworldlabs @gen_intuition @moonlake @physical_int @bifrost_ai @reactorworld @odysseyml @NianticSpatial @Figure_robot @SkildAI @DecartAI @runwayml @wayve_ai maybe add @rumik_ai in that list too 👀

here are some cool companies doing stuff in world models and physical ai - anyone im missing?? @amilabs @theworldlabs @gen_intuition @moonlake @physical_int @bifrost_ai @reactorworld @odysseyml @NianticSpatial @Figure_robot @SkildAI @DecartAI @runwayml @wayve_ai @agilityrobotics

@kadirnardev crwazy

The loss value of my new omni model is 1.5 👀

@kadirnardev exactly

Why can't teams with big resources achieve good audio quality? The model is nice but the audio quality is low :/

meet priya. and for the next 47 secs listen to her story.

@varundeepsaini Dayum brooo

This is either the biggest fumble or the biggest bag

eid shopping done

New blog!! From Air Pressure to Speech: Exploring classical TTS pipeline i walk you through how tts works from capturing sound froom microphone to getting wav files hope you guys enjoy ; ) @VioSIlverP/rkOZITb9Wx?utm_source=profile-share" target="_blank" rel="nofollow noopener">hackmd.io/@VioSIlverP/rk…

@AcousimHss no wonder where rest of the coffee went

My yearly perfmaxxing post Ice cream was still better than this coffee tho 😛

put out a new worklog on optimizing snake1d activation kernel in triton