Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Commits on main
NVFP4 QAT: accept fp16 autocast and non-contiguous activations (#14780)
d08356c Daniel Han committed 27 days ago
XPU: train an explicit bf16 load in bf16 under the fp16 idiom, and match vLLM's dtype when it refuses bf16 (#14606)
cdd0900 Daniel Han committed 27 days ago
RoPE: 64-bit cos / sin row offsets for tables past 2**31 elements (#14714)
37eec0a Daniel Han committed 27 days ago
Studio diffusion: cap the single-file host RAM refusal reserve at the fp8-storage reserve (#14758)
adf2a6c Daniel Han committed 27 days ago
Studio: keep one GGUF row per build in repos whose model name contains NVFP4 (#14733)
9d45fad Daniel Han committed 27 days ago
Honour gradient_checkpointing assigned on the config after construction (#15200)
0cbcf8d Daniel Han committed 27 days ago
Studio: keep LTX-2.5's bf16 text encoder on an unset request (#14737)
3422056 Daniel Han committed 27 days ago
Studio: keep MiniMax-H3 PDD float64 math on the host on a device without float64 (#14848)
0e9dd10 Daniel Han committed 27 days ago
RMSNorm / LayerNorm: do not hash a symbolic width in the launch settings cache (#14727)
da00729 Daniel Han committed 27 days ago
Generation keeps inference_mode on DeepSpeed ZeRO-1/2; only ZeRO-3 needs no_grad (#14673)
d4a92bb Daniel Han committed 27 days ago
Studio diffusion: pin streamed image text encoders with the background pinner, and the prefix that fits (#15206)
43798e2 Daniel Han committed 27 days ago
Studio: Mask from text runs SAM 3.1 in bf16 unless the card is short of memory (#14736)
9cd247d Daniel Han committed 27 days ago
MoE grouped GEMM autotune: key, tune and fit tiles on the experts' own GPU (#14719)
bd35fcc Daniel Han committed 27 days ago
Studio sd.cpp: a lost Vulkan device is not a strike; Arc / ROCm keep Vulkan; a UHD iGPU beside a card no longer vetoes Vulkan (#14610)
f4a21e2 Daniel Han committed 27 days ago
Prefetch: keep subfolder weights that a root safetensors index maps to (#15184)
3fe735b Daniel Han committed 27 days ago
gpt-oss: fix the unreadable lone .bin advice and pin the layout probe order to transformers' resolver (#14763)
12eb64b Daniel Han committed 27 days ago
Linear input-cast hook: skip float8-weight Linears (#14690)
2faad9f Daniel Han committed 27 days ago
Studio: judge a weight name's .. where the filesystem sends it, not by its text (#14800)
1610b38 Daniel Han committed 27 days ago
Studio: keep the AMD gfx target and HIP id map until hardware re-detection (#14970)
de03e06 Daniel Han committed 27 days ago
head_dim 512 attention: leave mask plus is_causal to SDPA, fall back on mask-scan OOM (#14691)
36f8f79 Daniel Han committed 27 days ago
Studio diffusion: price the compiled tier's requests from its own learned peak (offloaded bf16 keeps its residency after compile) (#15198)
c5922a0 Daniel Han committed 27 days ago
Cross entropy: an all -inf vocab chunk no longer turns the loss into NaN (#14713)
c1df811 Daniel Han committed 27 days ago
Installers: read lspci and the audio.cpp driver version once per run, one WMI query for both AMD scans (#14984)
1939a80 Daniel Han committed 27 days ago
Padding choice: per compute capability thresholds, RTX PRO 6000 / RTX 5090 peaks (#15216)
fd6d9f6 Daniel Han committed 27 days ago
Studio GGUF placement: the dGPU wins Auto only without losing context; explicit picks keep host-only bytes off the dGPU (#14615)
69d6ab9 Daniel Han committed 27 days ago
Cross entropy: correct gradients under torch.compile for vocabs above 65536 (#14712)
af89fcb Daniel Han committed 27 days ago
Studio: trim masked padding from pre-tokenized rows and range-check labels (#14818)
41db865 Daniel Han committed 27 days ago
DeepSeek-V4.1 context scratch: charge the dense KQ scores when flash attention is off (#14838)
54e09f8 Daniel Han committed 27 days ago
Tests: run the DDP no_sync accumulation steps' forward inside no_sync (#14711)
4647d61 Daniel Han committed 27 days ago
Studio: trust forwarded headers only behind an opted-in proxy or the managed tunnel (#14753)
f040073 Daniel Han committed 27 days ago