-
Notifications
You must be signed in to change notification settings - Fork 83
Pull requests: Avarok-Cybersecurity/atlas
Author
Label
Projects
Milestones
Reviews
Assignee
Sort
Pull requests list
feat(server): use shared model-agnostic canonical request/response types
#307
opened Jul 13, 2026 by
sercand
Loading…
6 tasks done
fix(loader): dequant NVFP4-packed q/k/v in qwen3_next attention (#299 follow-on)
#306
opened Jul 13, 2026 by
rsafier
Collaborator
Loading…
feat(dflash): policy layer: adaptive speculation + spec-through-think (stacked on #303)
#304
opened Jul 12, 2026 by
rrstesiak
Contributor
Loading…
3 of 4 tasks
feat(dflash): K=gamma base engine: paged drafter, fused decode+verify, verify-step kernels (stacked on #302)
#303
opened Jul 12, 2026 by
rrstesiak
Contributor
Loading…
3 of 4 tasks
feat(dflash): scheduler/server verify-step glue — raw-argmax bypass, T>0 stochastic accept, EAGLE ctx-fix default-on
#297
opened Jul 11, 2026 by
Sujimoshi
Contributor
Loading…
5 tasks done
perf(gdn-prefill): ldmatrix A+B GDN-projection GEMM — 2.1x kernel, warm-TTFT median -5.8% / p90 -10.3%
#296
opened Jul 10, 2026 by
tbraun96
Contributor
Loading…
feat(dflash): drafter head core — forward_block split, NVFP4-safe lm_head, γ-1 noise0-skip fix
#294
opened Jul 10, 2026 by
Sujimoshi
Contributor
Loading…
5 tasks done
feat(dflash): K=γ verify model orchestration + EAGLE ctx-accumulator contiguity fix
#292
opened Jul 10, 2026 by
Sujimoshi
Contributor
Loading…
5 tasks done
feat(dflash): SSM-layer K=γ verify kernels — WY16 GDN, batched phase-1 conv, opt-in F32 GDN output
#290
opened Jul 10, 2026 by
Sujimoshi
Contributor
Loading…
1 of 3 tasks
refactor(loader): gate batched-prefill MLA exclusion on capability, not model_type
#285
opened Jul 10, 2026 by
rsafier
Collaborator
Loading…
feat(gdn-tp): dense-model GDN/SSM TP-shard (dense pure-TP) [stacks on #254+#250]
#284
opened Jul 10, 2026 by
rsafier
Collaborator
Loading…
fix(toolcall): harden qwen3_coder tool calls against long-context echo attractor
#283
opened Jul 9, 2026 by
AzeezIsh
Collaborator
Loading…
feat(nemotron): bring up Nemotron-H Puzzle 75B-A9B (heterogeneous MoE + state=96 fix)
#281
opened Jul 9, 2026 by
AzeezIsh
Collaborator
Loading…
4 of 6 tasks
chore(deps): Bump io-uring from 0.7.12 to 0.7.13 in the patch-updates group across 1 directory
#273
opened Jul 6, 2026 by
dependabot
Bot
Loading…
fix(sampling): preset fallback (not greedy 0) when generation_config absent — #228 review guard
#267
opened Jul 5, 2026 by
rsafier
Collaborator
Loading…
perf(decode): n>=4 tiled QKV + o_proj GEMV for NVFP4 batched decode (extract of #214 onto main)
#266
opened Jul 5, 2026 by
rsafier
Collaborator
Loading…
test(harness): official single-GPU suite + soak + vision fixtures (recut of #212 onto main)
#265
opened Jul 5, 2026 by
rsafier
Collaborator
Loading…
feat(tokenizer): source chat templates from model checkpoint (recut of #221 onto main)
#264
opened Jul 5, 2026 by
rsafier
Collaborator
Loading…
fix(loader): gate CUTLASS grouped NVFP4 MoE on architecture, not model_type string
#260
opened Jul 5, 2026 by
rsafier
Collaborator
Loading…
feat(dflash): model glue/scaffolding for K=γ verify + native NVFP4 lm_head
#258
opened Jul 5, 2026 by
Sujimoshi
Contributor
Loading…
4 tasks done
perf(spark-runtime): opt-in NVTX kernel-attribution + phase-timing instrumentation
#255
opened Jul 5, 2026 by
Sujimoshi
Contributor
Loading…
3 of 4 tasks
Previous Next
ProTip!
Type g i on any issue or pull request to go back to the issue listing page.