All issues

Search results

SPEC-DSPARK: dense 27B lane accepts 12.2% where the pinned oracle accepts 49.3% (cross-engine 0.35x)

Status: Open.

#430 In mudler/vllm.cpp;

· localai-bot opened on Aug 11, 2026

Ignored numeric/array bounds make the grammar a superset, not a subset -- and remove the only termination guard

Status: Open.

#429 In mudler/vllm.cpp;

· filipsajdak opened on Aug 11, 2026

test_serve_low_tools: the concurrency-cap assertion races the server-side counter and reads 3 under load

Status: Open.

#428 In mudler/vllm.cpp;

· localai-bot opened on Aug 11, 2026

Six of the eight activation-width resolvers have INTERNAL linkage in qwen3_5.cpp — a new model literally cannot inherit the dtype decision, so it writes a literal (1255 of them)

Status: Open.

#426 In mudler/vllm.cpp;

· localai-bot opened on Aug 11, 2026

CORRECTNESS: Kimi-Linear MoE scatter sizes row_bytes from the STREAM dtype but its source/dest are hardcoded f32 — VT_KIMI_BF16_STREAM=1 copies half a row at half stride, and it confounds a recorded negative result

Status: Open.

#423 In mudler/vllm.cpp;

· localai-bot opened on Aug 11, 2026

ENG-TRAILER-MERGE-ARTIFACTS closed DONE on main without updating docs/STATUS.md or docs/BENCHMARKS.md

Status: Open.

#420 In mudler/vllm.cpp;

· localai-bot opened on Aug 11, 2026

The forge's Co-authored-by line is attribution, not an authorship claim: bot-opened PRs red main on merge

Status: Open.

#418 In mudler/vllm.cpp;

· localai-bot opened on Aug 11, 2026

We emit F32 where vLLM emits bf16 on the biggest prefill buffers (GDN in_proj, attn gate, q/k/v) — 2x traffic, per-token, flat across concurrency; and the canonical harness ENFORCES the 0.25.0 rollback oracle

Status: Open.

#417 In mudler/vllm.cpp;

· localai-bot opened on Aug 11, 2026

THE PREFILL GAP IS PROBABLY THE KV DTYPE: the oracle runs kv_cache_dtype=fp8_e4m3 (confirmed at runtime), we store bf16 — it moves HALF the KV bytes we do

Status: Open.

#415 In mudler/vllm.cpp;

· localai-bot opened on Aug 11, 2026

Canonical grids measure vLLM WITHOUT --language-model-only: the oracle runs the UNFUSED QK-norm+RoPE+gate path while we run the fused one — our published ratios are FLATTERED

Status: Open.

#414 In mudler/vllm.cpp;

· localai-bot opened on Aug 11, 2026

No Muse Glimmer image preprocessor: callers pick their own grid, silently changing the token count

Status: Open.

#413 In mudler/vllm.cpp;

· localai-bot opened on Aug 11, 2026

12 of 54 tests/scripts suites are executed by nothing, and check-test-registration.py cannot see the class

Status: Open.

#408 In mudler/vllm.cpp;

· localai-bot opened on Aug 11, 2026