rajveer43

Issues somebody may already be working on are listed last. List everything by date

• SnapKVOptions silently turns snap_batched_scoring off, and an explicit flat kwarg equal to its default loses to options

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 3/5 1-2 days Newbie friendliness 68/100

  rajveer43/VeloxQuant-MLX#662 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• PrefixCache NOT_TRIMMABLE notice is false: exact full-prompt repeats are not reused for h2o/snapkv/streaming_llm

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 3/5 1-2 days Newbie friendliness 76/100

  rajveer43/VeloxQuant-MLX#661 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• benchmarks/model_kv_benchmark.py reports fp16 perplexity for every method and miscounts latency tokens

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 3/5 1-2 days Newbie friendliness 74/100

  rajveer43/VeloxQuant-MLX#660 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• Artifact stores save codebooks as fp16, so a cold store and a warm store give different encodings

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 3/5 1-2 days Newbie friendliness 70/100

  rajveer43/VeloxQuant-MLX#659 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• dp_allocate_bits gives different (suboptimal) allocations when all variances are scaled down

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 3/5 1-2 days Newbie friendliness 72/100

  rajveer43/VeloxQuant-MLX#658 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• RecursivePolarTransform folds an angle of exactly π/2 to 0 at levels ≥ 2

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 2/5 1-3 hours Newbie friendliness 82/100

  rajveer43/VeloxQuant-MLX#657 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• PooledKVCache.reset() / release() leave len() at the old token count

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 2/5 1-3 hours Newbie friendliness 88/100

  rajveer43/VeloxQuant-MLX#656 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• _tova_update_batched deferred path returns values in the input dtype instead of fp16

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 2/5 1-3 hours Newbie friendliness 77/100

  rajveer43/VeloxQuant-MLX#655 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• chunkkv_update crashes with a raw MLX gather error when n_sink=0 and chunk_size > budget; the batched version silently empties the cache

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 2/5 1-3 hours Newbie friendliness 78/100

  rajveer43/VeloxQuant-MLX#654 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• NSNQuant codebook module cache is keyed without n_samples or iters, so a cheap codebook is returned for the default request

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 2/5 1-3 hours Newbie friendliness 88/100

  rajveer43/VeloxQuant-MLX#653 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• H2O/Keyformer fused-evict 'bit-for-bit equivalent' claim is false (1-ULP fp16 key differences vs the MLX twin)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 3/5 1-2 days Newbie friendliness 58/100

  rajveer43/VeloxQuant-MLX#652 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• vecinfer_encode_decode_metal uses smooth row h % smooth_rows while the MLX twin averages rows when smooth_rows != H: backends disagree

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 3/5 1-2 days Newbie friendliness 72/100

  rajveer43/VeloxQuant-MLX#651 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• h2o_fused_evict / keyformer_fused_evict with no evictable row (all protected, or NaN scores) use index −1: out-of-bounds read and corrupted positions/keys

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 4/5 3-5 days Newbie friendliness 52/100

  rajveer43/VeloxQuant-MLX#650 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• Q-Filters Metal fused eviction keeps different rows from its MLX twin on score ties, and stores +inf scores for protected rows

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue

  Difficulty 4/5 3-5 days Newbie friendliness 68/100

  rajveer43/VeloxQuant-MLX#649 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• Planner _retention undercounts kept tokens for hybrid (quant + eviction) methods

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 2/5 1-3 hours Newbie friendliness 86/100

  rajveer43/VeloxQuant-MLX#648 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• MLXBlockStorage.resident_bytes() under-reports int4/int2/int1 blocks by 2×/4×/8× (buffers are unpacked uint8)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 2/5 1-3 hours Newbie friendliness 86/100

  rajveer43/VeloxQuant-MLX#647 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• CacheRoute admission charges the uncapped shard count, so the hottest session is rejected while a cold one is admitted

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 2/5 1-3 hours Newbie friendliness 76/100

  rajveer43/VeloxQuant-MLX#646 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• CacheRoute RateEstimator does not measure rate: every owner saturates near 29.4 and silent owners never decay

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 3/5 1-2 days Newbie friendliness 67/100

  rajveer43/VeloxQuant-MLX#645 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• Spectral rotation disk cache is keyed only on model_name (default "model"), so a second model silently gets the first model's rotations

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 4/5 3-5 days Newbie friendliness 53/100

  rajveer43/VeloxQuant-MLX#644 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• SpectralQuant calibration only samples KV head 0

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 3/5 1-2 days Newbie friendliness 74/100

  rajveer43/VeloxQuant-MLX#643 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• CrossModelMapper.load_layer indexes layers by list position, not target layer id

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 2/5 1-3 hours Newbie friendliness 86/100

  rajveer43/VeloxQuant-MLX#642 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• zipcache_bytes(state) ignores state.group_size and assumes 32

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 2/5 1-3 hours Newbie friendliness 86/100

  rajveer43/VeloxQuant-MLX#641 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• RaBitQQuantizer.compression_ratio reports 16× but actual storage gives 9.14× (metadata not counted)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 2/5 1-3 hours Newbie friendliness 88/100

  rajveer43/VeloxQuant-MLX#640 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• BenchmarkDatabase feeds the planner false 'measured' data: chip missing from fingerprint, memory_reduction defaults to 1.0, corrupt placeholders still match

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 4/5 3-5 days Newbie friendliness 55/100

  rajveer43/VeloxQuant-MLX#639 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• Mac recommender's KV-size estimate divides full K+V by a key-only ratio and ignores streaming_llm's fixed window

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 2/5 1-3 hours Newbie friendliness 84/100

  rajveer43/VeloxQuant-MLX#638 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• veloxquant profile --set crashes on array fields (stale copy of serve's parse_overrides)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 2/5 1-3 hours Newbie friendliness 82/100

  rajveer43/VeloxQuant-MLX#637 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• OutlierSplitHandler cannot run: encode crashes on mx.array input, decode sizes its output from the inlier slice

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 2/5 1-3 hours Newbie friendliness 78/100

  rajveer43/VeloxQuant-MLX#636 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• Control-panel log pane stops updating permanently after 500 lines

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 3/5 1-2 days Newbie friendliness 84/100

  rajveer43/VeloxQuant-MLX#635 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• NpyArtifactStore._atomic_save temp names collide across processes: concurrent writers crash and readers can see a truncated .npy

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 3/5 1-2 days Newbie friendliness 76/100

  rajveer43/VeloxQuant-MLX#634 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• Observers never receive events: KVCacheConfig.observers / with_observer() are never read, and bit_width_outlier is ignored

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 4/5 3-5 days Newbie friendliness 48/100

  rajveer43/VeloxQuant-MLX#633 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• veloxquant recommend (legacy mode) recommends methods that cannot be built or served (rabitq, spectral)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 3/5 1-2 days Newbie friendliness 76/100

  rajveer43/VeloxQuant-MLX#632 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• WorkloadSpec.fp16_kv_bytes() omits the KV-head count, so select_kv_cache_config's memory-pressure rule effectively never fires

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 4/5 3-5 days Newbie friendliness 68/100

  rajveer43/VeloxQuant-MLX#631 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• HardwareProfiler 'available memory' is physical RAM minus this process's MLX usage: planner recommends a 22.9 GiB KV cache on a 24 GB M4

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 3/5 1-2 days Newbie friendliness 75/100

  rajveer43/VeloxQuant-MLX#630 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• ProfileReport.total_tokens and tokens_per_sec count every token once per layer (24× too high on Qwen2.5-0.5B)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 3/5 1-2 days Newbie friendliness 84/100

  rajveer43/VeloxQuant-MLX#629 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• Profiler compression ratio uses a key-only, single-head fp16 baseline: real compression is reported as expansion

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 3/5 1-2 days Newbie friendliness 78/100

  rajveer43/VeloxQuant-MLX#628 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• veloxquant estimate-memory --model-config ignores the config's geometry (baseline 10.7× too large)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 2/5 1-3 hours Newbie friendliness 88/100

  rajveer43/VeloxQuant-MLX#627 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• AMC's MID (8-bit) and LOW (4-bit) tiers never quantize (lossless), while bytes are charged at 8 and 4 bits

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 3/5 1-2 days Newbie friendliness 76/100

  rajveer43/VeloxQuant-MLX#626 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• TurboQuantRVQ default residual_scale treats an MSE bound as a standard deviation: stage-2 codebook is 1.6–7.7× too narrow (default method)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 2/5 1-3 hours Newbie friendliness 86/100

  rajveer43/VeloxQuant-MLX#625 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• flash_prefill_attend / rabitq_prefill_attend accumulate Q·Kᵀ in half: error is 100–350× MLX's fp16 SDPA at realistic logit scales

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 3/5 1-2 days Newbie friendliness 74/100

  rajveer43/VeloxQuant-MLX#624 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• trim() on KIVI, KIVI-sink, NSNQuant, Kitty and AdaKV doesn't roll back the quantization frontier / running stats: rewritten rows stay fp16 but are counted as compressed

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 4/5 3-5 days Newbie friendliness 52/100

  rajveer43/VeloxQuant-MLX#623 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• 32 of 38 cache methods (and VecInfer's Metal path) return fp16 for bf16 input: values above 65504 become inf/NaN and attention is promoted to fp32

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P2

  Difficulty 4/5 3-5 days Newbie friendliness 48/100

  rajveer43/VeloxQuant-MLX#622 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• uint8 index overflow for codebooks with more than 256 entries (CommVQ b > 8, NSNQuant nsn_codebook_size > 256)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P1

  Difficulty 3/5 1-2 days Newbie friendliness 68/100

  rajveer43/VeloxQuant-MLX#621 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• Planner memory model credits phantom savings: default recommendation (zipcache, "99.4% saved") keeps 100% of the fp16 cache resident (#606 pattern)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P1

  Difficulty 4/5 3-5 days Newbie friendliness 55/100

  rajveer43/VeloxQuant-MLX#620 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• turboquant_fused_rvq_decode_attend returns garbage for D > 256 and for GQA inputs, with no validation (register array overflow, #603 pattern)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P1

  Difficulty 4/5 3-5 days Newbie friendliness 52/100

  rajveer43/VeloxQuant-MLX#619 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• transfer: fit_mapper stores list positions as source-layer ids, so transfer_cache silently maps from the wrong layers

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P1

  Difficulty 3/5 1-2 days Newbie friendliness 74/100

  rajveer43/VeloxQuant-MLX#618 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• SpectralQuantizer scales inputs to [-1, 1] but quantizes with a codebook spanning only ±0.19, so most coordinates clip

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P1

  Difficulty 4/5 3-5 days Newbie friendliness 55/100

  rajveer43/VeloxQuant-MLX#617 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• VecInfer dual transform breaks under GQA with per-KV-head smooth factors: queries get the head-mean, keys get per-head factors

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P1

  Difficulty 3/5 1-2 days Newbie friendliness 74/100

  rajveer43/VeloxQuant-MLX#616 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• GEAR quantize_base axes are swapped: keys get per-token instead of per-channel quantization (18× the error with an outlier channel)

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P1

  Difficulty 3/5 1-2 days Newbie friendliness 86/100

  rajveer43/VeloxQuant-MLX#615 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• A2ATS reports is_trimmable() == True, but trim() leaves _next_position stale, so the near/far RoPE split targets a future position

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P1

  Difficulty 4/5 3-5 days Newbie friendliness 48/100

  rajveer43/VeloxQuant-MLX#614 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day

• A2ATS applies RoPE a second time to keys that mlx_lm has already rotated

  Open

  ⋯Report a problem with this issue

  • Is something wrong?

  • Already taken
  • Not beginner-friendly
  • Stale or invalid
  • Something else…
  • Send Cancel
  bug bughunt good first issue P1

  Difficulty 4/5 3-5 days Newbie friendliness 52/100

  rajveer43/VeloxQuant-MLX#613 · Opened Sep 30, 2026

  Maintainers usually reply within 1 day