Spaces:
Running
Running
Commit History
19 days ago
-
Composer final; probes off Running
Verified signature15a87ca
20 days ago
-
Composer fixes; clean 4-bit probe
Verified signatured335996 -
Quantization Composer, portable save, KV quant, OpenVINO/AutoRound/SINQ probe
Verified signatureee6a75e -
4-bit backend probe
Verified signaturec3e0722 -
Explicit Load button; env-gated llama.cpp probe
Verified signature1eca699 -
one logical CPU per core, compressed-tensors float32 on CPU
Verified signature04d4603 -
THP allocator, core-pinning probe
Verified signature0c38057 -
bandwidth/hugepage/placement probe, forward-share log, mask-builder kwarg filter
Verified signaturec962244 -
v4 mask-kwarg shim (Spark-X), speculators anchor-sampling detection, CPU draft-length advice
Verified signature0f1956b -
normalized-aux drafters (T5Gemma), quantized drafter matmuls, visible generate retries, fp16 speculation advice
Verified signature8922fe1 -
dynamic_fp16 scheme, eager attention for eager-only remote code, enc-dec drafter start token, quantized draft models, spec timing
Verified signaturef9ed677 -
Speculative: deferred commit for recurrent caches, enc-dec decoder discovery + width check, cumulative aux calibration, variant configs; v4 tied-weights shim; gated/native config guards; loader attempt logging
Verified signature117a403 -
Speculative: version-tolerant KV crop, real layer-skip probe (block output kinds), repo::variant drafters
Verified signaturecce4354 -
Dynamic int8 in place, one Linear at a time (no deepcopy / full fp32 copy: fixes OOM on 2B models); compile-stable int4/int8 kernels
Verified signature7ca9e8c -
Release compiled forward + block drafters on model eviction; raw-logit token probabilities; README
Verified signature1c280ce -
QwenScope instruments (two-text comparison, positions, strength presets, original vs steered), speculative decoding, torch.compile, native int4/int8 kernels, sources
Verified signature902a1a8 -
Speculative decoding (draft, self, DFlash family), torch.compile, int8/int4 kernels
Verified signature703682d -
Hardware probe, native int8/int4 CPU kernels
Verified signaturedba7466 -
Log CPU model and native bf16 on a separate INIT line
Verified signature5b43e68 -
Decide auto dtype from native bf16 flags; log CPU model
Verified signature970e1c9 -
CPU decode: fp32 auto when bf16 is slower; quanto int8 on bf16 activations; flag CPU float8
Verified signature0b41e30
22 days ago
-
DFT/ALSO: setup lines to Space log, final row-weight summary, no perplexity for DFT
Verified signature7218d58 -
DFT training method + native ALSO optimizer (per-row DRO weights)
Verified signature03b83f2 -
Update Small LM Bistro
Verified signaturee762c2d -
SAR: drop stale prompts from other behaviours; manifold steering card warns when the layer does not drive the answer
Verified signature5595019 -
MANIFOLDS tab, SAR, revised SAE training, NLA reading discipline, tuning advisor, sources
Verified signaturec1edad6
23 days ago
-
Rename NLA tab to NLA · SHOW & TELL
Verified signaturedcafe31
24 days ago
-
Hover hints on checkbox labels
Verified signature3ef05f8 -
Oracle records note attached steering; README
Verified signature3a84f79 -
Activation Oracle reader in NLA · TALKING CURE
Verified signatureb2edced -
README: SAE · COANALYSIS, NLA · TALKING CURE, steering, NanoQuant, GPart/GAWS
Verified signature249bdb9 -
Steering panel: read contrasts from the unsteered model
Verified signaturea5bf697 -
Inference-tab activation steering panel (CAA/ActAdd, directional ablation, α sweep); sink-safe contrasts; hints on file labels
Verified signaturec3dc21c -
NLA · TALKING CURE tab (Nano-NLA / kitft NLAs: read, scan, model diff, edit steering); SAE tab renamed SAE · COANALYSIS
Verified signature85d4299 -
Lab: SAE-COANALYST tab, NanoQuant, GPart/GAWS, steering registry
Verified signature2eacce3
25 days ago
-
Update README.md
Verified signaturecad5d38
26 days ago
-
Auto training precision: keep bf16 for models whose fp32 weights would exceed 1/3 of RAM
Verified signature061fd38 -
Atelier audit: fix DPO ref-free/losses, CPT chunking, Decision scoring+calibration, Vision SFT, 8/4-bit AdamW, bespoke-model training, hot-swap; chat template presets; Hub model/dataset links
Verified signaturef1f0998 -
Rename tab to ATELIER · TRAINING
Verified signatureb3884b3 -
Support Universal programmatic architecture synthesis, on-the-fly safetensors conversion, and custom in-repo architectures (PoetLM, OpenPeerLLM, Aurora1.0, Aurora-80K, Axiom-1, circuit-sparsity)
Verified signatureb8c149e
27 days ago
-
Support openai/circuit-sparsity and prune unlaunchable non-Transformers models (bistro/engine.py)
Verified signaturedb4cf67 -
Support openai/circuit-sparsity and prune unlaunchable non-Transformers models (bistro/arch_bridge.py)
Verified signaturef647a85 -
Support openai/circuit-sparsity and prune unlaunchable non-Transformers models (bistro/config.py)
Verified signatureabab780 -
Quantized checkpoints saved as safetensors (+legacy .pt/.bin loading); architecture bridge (native retry, in-repo code discovery, clear errors)
Verified signature4de6884 -
Private-repo adapter loading: HF token fields, hot-swap from Hub, checkpoint subfolder support
Verified signature451f248 -
Fix: native Muse2 support, Mamba/hybrid LoRA target compat (Falcon-H1), Granite-hybrid num_experts config, clearer load errors
Verified signatureeedcbc7 -
Make message entry full-width above action buttons & optimize mobile layout
Verified signaturea151e61
28 days ago
-
Fix base/custom model inference (Pragya-EXP2, Supra2-100M, cagliostro), CPU LoRA meta-tensor/FlexAttention issues & add sub-banner note
Verified signaturef23626a -
Merge branch #AlekseyCalvin/Small_LM_Bistro' into 'AlekseyCalvin/SLM_Bistro'
50698e3 -
Update README.md
Verified signature224b982