Hugging Face's logo
Dipankar Sarkar's picture
🏗️ Building on HF

Dipankar Sarkar PRO

dipankarsarkar
3 249 17
John6666's profile picture ucr-max's profile picture WANG-CC's profile picture
·
https://www.dipankar.cc

AI & ML interests

Building the AI-native stack. Agents as infrastructure, safety as architecture, performance as plumbing. I publish the receipts: papers, datasets, demos.

Recent Activity

repliedto ManniX-ITA's post about 3 hours ago
# opencoti-llamafile 0.10.3-c5 — settled admission for multi-agent serving New cut of the opencoti single-file inference engine (llamafile 0.10.3 / llama.cpp + 87 additive patches). Zero-dependency APE: one executable for Linux, Windows, macOS & BSD. What's new vs c4: **PolyKV fan-out — pool from a live session.** `POST /polykv/pools` gains `from_session`/`from_slot`: the shared prefix is snapshotted server-side from the session's cached KV — no tokens resent, token-exact. Ephemeral pools auto-release when orchestrators die mid-round. **PolyKV P7 — settled admission.** Spawning agents faster than the tps signal settles was oversubscribing pools. Now: a per-pool settle window paces admits just enough for a reliable reading; warming sessions no longer bias the mean; the post-admit forecast uses the measured per-admit drop; idle gaps (agents mid-tool-call) no longer read as free capacity; `guarantee_min_sessions` means a new/nested pool always gets its first agent — capacity checks can never deadlock an orchestrator; the enforced gate applies to new sessions only, with per-request `overcommit`. Benchmark (multi-agent courier, floor 15 tok/s): time-under-floor −52%, deep sub-floor −83%, delivery p50 −43%, 100% task score. **Zero-conf GPU sharing.** Instances on one GPU discover each other over shared memory — no ports, no config — and split compute by `--gpu-share-weight`. Measured (3090): weights 2:1 → 71.7/36.2 tok/s; holds at `--parallel 4` and under MTP. Idle peers cost nothing (solo = full speed), crashes age out in 3 s; `GET /gpu/peers` shows live shares + busy %. From c5 every release ships per-platform side-load DSOs: `dso/<ver>/` with Linux x86_64 + sbsa `.so` and a Windows `.dll`. https://huggingface.co/ManniX-ITA/opencoti-llamafile
liked a model about 4 hours ago
owensong/Inflect-Nano-v2
repliedto owensong's post about 4 hours ago
I just released Inflect v2, pushing complete TTS below 4M parameters. Inflect-Nano-v2 fits every learned inference component, including waveform generation, into 3.96M parameters (15.97 MB FP32), making it one of the smallest publicly available complete TTS models. Inflect-Micro-v2 stays under 10M at 9.36M parameters (37.53 MB), using the extra capacity for clearer, more natural speech. This was not about building a toy that technically produces audio. I wanted to see how much useful TTS could actually fit at this scale. Nano stayed surprisingly close to Micro across the reported release metrics, but if you want to hear Inflect at its best, start with Micro: https://huggingface.co/owensong/Inflect-Micro-v2 Text goes in and 24 kHz speech comes out without a second learned model. Both models have PyTorch and ONNX releases: PyTorch runs on CPU or CUDA, while the browser demo runs ONNX locally through WebGPU with a WASM fallback. The Inflect code and weights are released under Apache 2.0. Try Nano and Micro side by side: https://huggingface.co/spaces/owensong/Inflect-v2 Some people may remember the much rougher Inflect-Nano-v1 experiment. V2 was rebuilt and retrained to address its biggest weaknesses: metallic audio, unstable timing, weak prosody, and difficulty with unseen text. Nano-v2 is smaller while sounding substantially cleaner and more stable. In an anonymous community blind test across nine voices, Micro and Nano placed second and third by normalized preference rate. Full results, samples, limitations, and inference instructions are on the model pages. Inflect-Nano-v2: https://huggingface.co/owensong/Inflect-Nano-v2 I built Inflect independently with a limited training budget. If you test it, give it something difficult: unusual names, numbers, abbreviations, symbols, or awkward punctuation. Send me the exact text when something breaks. If you build something with it, I would genuinely love to see it.
View all activity

Organizations

Skelf Research's profile picture Neul Labs's profile picture Cognisoc's profile picture Incredlabs's profile picture