All issues

Search results

feat: configurable DeltaNet state-checkpoint floor (and q4/q8 KV) — 24 GB cards can't serve long context even when it's ~1 GiB away from fitting

enhancementNew feature or requestNew feature or request

Status: Open.

#33 In truespar/paddock;

· orangeswim opened on Sep 26, 2026

Studio model-maximum reply budgeting overrides output ceilings with a 512-token floor

Status: Open.

#24 In truespar/paddock;

· DivyamTalwar opened on Sep 18, 2026

Qwen35: incremental two-GPU tensor splitting with parity and capacity gates

Status: Open.

#22 In truespar/paddock;

· ErikBPF opened on Sep 13, 2026

GPU access: Nsight Compute on the i-quant dense lanes, and the Flash Next GGUF lane on a card that seats the experts

Status: Open.

#20 In truespar/paddock;

· L4GN opened on Sep 5, 2026

Allow download and running of other quantizations apart from q4

enhancementNew feature or requestNew feature or request

Status: Open.

#13 In truespar/paddock;

· jezztify opened on Sep 5, 2026

e4m3 lm_head lane (pd_f8_gemm_lin_kernel<32,2,4>) faults with a misaligned shared address on sm_120

Status: Open.

#6 In truespar/paddock;

· L4GN opened on Sep 4, 2026

fp8 paged partial attention hd=128 splits=3 diverges 4.4e-3 from dense on sm_120

Status: Open.

#5 In truespar/paddock;

· L4GN opened on Sep 4, 2026