Search results

Speculative decoding (draft-dspark) silently falls back to CPU on DGX Spark / unified-memory CUDA devices — 2.35x slower than fixed

Status: Open.

#112 In PrismML-Eng/Bonsai-demo;

· thadreber-web opened on Jul 18, 2026

DSpark drafter: measured 32k enablement ceiling + net-negative speedup at depth (mechanism ≠ VRAM)

Status: Open.

#105 In PrismML-Eng/Bonsai-demo;

· cnndabbler opened on Jul 17, 2026

Together serverless Ternary Bonsai 27B endpoint metadata points at an AWQ-4bit artifact, catalog says ternary g128

Status: Open.

#102 In PrismML-Eng/Bonsai-demo;

· Astezelex opened on Jul 17, 2026

Any plans to release the ternary vLLM kernels used in the whitepaper evals?

Status: Open.

#101 In PrismML-Eng/Bonsai-demo;

· Astezelex opened on Jul 17, 2026

VAE decode OOMs above 256x256 on WebGPU adapters without shader-f16 (Pascal & older) - found the root cause, made a working fix

Status: Open.

#96 In PrismML-Eng/Bonsai-demo;

· ArtyomITA opened on Jul 16, 2026

Question: Exact PrismML-MLX commit used for Bonsai-27B-mlx-1bit release

Status: Open.

#95 In PrismML-Eng/Bonsai-demo;

· twinfufu opened on Jul 16, 2026

Pre-built llama.cpp binaries: Metal fails to compile on macOS 26 Tahoe (M5)

Status: Open.

#93 In PrismML-Eng/Bonsai-demo;

· lightoshadow opened on Jul 16, 2026

Ternary-Bonsai-27B on ROCm / RDNA3: works unpatched (RX 7800 XT, gfx1101) + benchmarks

Status: Open.

#92 In PrismML-Eng/Bonsai-demo;

· base-zz opened on Jul 15, 2026

Ornith-1.0 9B version?

Status: Open.

#86 In PrismML-Eng/Bonsai-demo;

· hrstoyanov opened on Jul 15, 2026

is iphone's NPU utilized in Bonsai inference?

Status: Open.

#85 In PrismML-Eng/Bonsai-demo;

· ColdCodeCool opened on Jul 15, 2026

Ternary-Bonsai-27B Q2 produces gibberish; Ternary-Bonsai-8B Q2 works on same host/build

Status: Open.

#82 In PrismML-Eng/Bonsai-demo;

· facundoem94 opened on Jul 14, 2026

GLM 5.2 version ?

Status: Open.

#81 In PrismML-Eng/Bonsai-demo;

· X-Ryl669 opened on Jul 14, 2026