Snapshot date: 2026-07-09 Total records: 350
This is the deterministic, navigable rendering of the canonical mlx-model-porting/assets/sources.yaml registry. The registry order is author-curated; this view sorts records by ID without rewriting that order.
Local workload receipts and their limitations are evaluated in the Benchmark assessment. A catalogue citation or source-reported number does not prove target-workload performance.
Review depth states how far a source was reviewed and used by this project. It is not an MLX reproduction status.
| Review depth | Records | Meaning |
|---|---|---|
indexed |
24 | Catalogued for future review; do not represent as fully reviewed. |
screened |
190 | Abstract/readme/code path reviewed for relevance and limits. |
synthesized |
136 | Directly informed a runbook, rule, or technique decision. |
Support scope states whose implementation or receipt the evidence covers. It never widens a source beyond its recorded snapshot.
| Support scope | Records | Meaning |
|---|---|---|
context_only |
0 | Context, discovery, or source-selection evidence. It does not establish implementation support or performance. |
local_reproduced |
0 | Local benchmark or validation evidence. It applies only to the captured hardware, software, model, inputs, and workload. |
not-classified |
327 | No support scope is encoded. The record is citation evidence only and grants no implicit support claim. |
official_mlx |
8 | Apple-maintained MLX framework evidence. It establishes source or API support at the recorded snapshot, not target-workload performance. |
official_mlx_project |
4 | Apple-maintained project evidence built on MLX. It establishes project-scoped behavior, not a framework-wide guarantee. |
paper_only |
1 | Algorithm or architecture evidence without a demonstrated MLX implementation. Reproduce the path and pass parity before recommending it. |
third_party_pinned |
10 | Pinned non-Apple MLX implementation evidence. Treat it as reproducible prior art, not an official framework guarantee. |
Every source is constrained by the following claim-boundary rules:
- Review depth describes how the source informed this skill; it is not proof that the source was reproduced on MLX.
- A source without evidence_class, support_scope, or claim_types is catalogue evidence only and grants no implicit implementation, performance, memory, or quality claim.
- Performance, memory, quality, and audio-quality claims stay attached to the exact recorded hardware, software, model, inputs, workload, and metric.
- A record note may narrow these boundaries. The narrowest applicable boundary always wins.
Claim types describe what a classified record may support; absent claim types grant no implicit claim.
| Claim type | Records | Meaning |
|---|---|---|
algorithm_candidate |
0 | The source supports considering an algorithm, not claiming that the repository implements it. |
api_support |
8 | The cited API exists at the recorded snapshot; compatibility with a target model still requires validation. |
audio_quality |
1 | An audio-quality claim requires the cited listening or objective evaluation conditions. |
memory |
1 | A memory claim is limited to the exact measured setup and accounting method. |
mlx_implementation |
15 | An MLX implementation exists in the cited scope; this does not by itself prove parity or speed on another target. |
performance |
3 | A latency or throughput claim is limited to the exact measured setup and must not be generalized. |
quality |
0 | A quality claim is limited to the recorded task, dataset, metric, and configuration. |
risk_or_negative |
1 | The source records a failure, limitation, or risk; it is not positive support for the technique. |
serving_semantics |
7 | The source supports stated scheduling, caching, or isolation behavior only in the cited implementation. |
| Kind | Count |
|---|---|
paper |
227 |
repository |
49 |
official-doc |
32 |
source-code |
26 |
release |
8 |
technical-blog |
7 |
issue-report |
1 |
| Evidence class | Count |
|---|---|
not-classified |
327 |
primary_source_code |
13 |
official_api_doc |
7 |
release_note |
2 |
primary_paper |
1 |
| Topic | Records |
|---|---|
mlx |
57 |
speculative-decoding |
37 |
tts |
35 |
quantization |
34 |
audio-codec |
25 |
serving |
21 |
vocoder |
18 |
asr |
16 |
streaming |
16 |
attention |
15 |
kv-quantization |
15 |
moe |
14 |
kernels |
11 |
multimodal |
10 |
vision-language |
10 |
audio |
9 |
decoder-transformer |
9 |
long-context |
9 |
prefix-cache |
9 |
diffusion |
8 |
cache |
7 |
continuous-batching |
7 |
metal |
7 |
mlx-audio |
7 |
sparse-attention |
7 |
agent-skills |
6 |
audio-language |
6 |
flow-matching |
6 |
kernel |
6 |
kv-compression |
6 |
on-device |
6 |
recurrent |
6 |
stt |
6 |
weight-quantization |
6 |
codec |
5 |
kv-cache |
5 |
kv-pruning |
5 |
mlx-lm |
5 |
mlx-vlm |
5 |
runtime |
5 |
video |
5 |
vllm-mlx |
5 |
audio-generation |
4 |
batching |
4 |
benchmarking |
4 |
encodec |
4 |
flow |
4 |
gan |
4 |
lora |
4 |
memory |
4 |
mixed-precision |
4 |
multi-token |
4 |
non-autoregressive |
4 |
rotation |
4 |
ssm |
4 |
training |
4 |
vlm |
4 |
activation-quantization |
3 |
audio-lm |
3 |
benchmark |
3 |
block-streaming |
3 |
custom-kernel |
3 |
custom-kernels |
3 |
grid-sample |
3 |
hugging-face |
3 |
llava |
3 |
low-bit-formats |
3 |
outliers |
3 |
package |
3 |
source-selection |
3 |
speech |
3 |
speech-encoder |
3 |
speech-to-speech |
3 |
survey |
3 |
transformer |
3 |
whisper |
3 |
apple-silicon |
2 |
architectures |
2 |
bitnet |
2 |
block-sparse |
2 |
codec-lm |
2 |
codec-tokens |
2 |
compile |
2 |
conformer |
2 |
contributors |
2 |
conversion |
2 |
coverage |
2 |
cuda |
2 |
encoder-decoder |
2 |
examples |
2 |
export |
2 |
fourier |
2 |
fsq |
2 |
fusion |
2 |
gather |
2 |
gather-qmm |
2 |
generation |
2 |
gqa |
2 |
hadamard |
2 |
hybrid |
2 |
language-models |
2 |
lazy-evaluation |
2 |
linear-attention |
2 |
model-card |
2 |
moss-tts |
2 |
musicgen |
2 |
normalization |
2 |
nvfp4 |
2 |
omni |
2 |
operators |
2 |
paged-attention |
2 |
paged-kv |
2 |
parallel-codebooks |
2 |
parity |
2 |
prefix-sharing |
2 |
quantized-matmul |
2 |
release |
2 |
rope |
2 |
safetensors |
2 |
scaling-laws |
2 |
source-oracle |
2 |
speech-tokenizer |
2 |
swift |
2 |
tests |
2 |
threading |
2 |
validation |
2 |
vision-cache |
2 |
visual-token-pruning |
2 |
visual-token-reduction |
2 |
weights |
2 |
agent-serving |
1 |
api |
1 |
async-eval |
1 |
attention-sinks |
1 |
audiogen |
1 |
autograd |
1 |
automatic-prefix-cache |
1 |
autoregressive |
1 |
bart |
1 |
cache-safety |
1 |
canary |
1 |
codex |
1 |
compute |
1 |
context-sharding |
1 |
convolution |
1 |
correctness |
1 |
correctness-fix |
1 |
cpp |
1 |
cross-attention |
1 |
custom-function |
1 |
dac |
1 |
delay-pattern |
1 |
dependencies |
1 |
development |
1 |
dflash |
1 |
dialogue |
1 |
diffusion-drafter |
1 |
diffusion-lm |
1 |
distillation |
1 |
distributed-attention |
1 |
distributed-inference |
1 |
dit |
1 |
downloads |
1 |
dual-track |
1 |
duplex |
1 |
duration |
1 |
eagle |
1 |
eagle2 |
1 |
edge |
1 |
encoder |
1 |
encoder-transformer |
1 |
eviction |
1 |
fastvlm |
1 |
format |
1 |
fp4 |
1 |
gather-mm |
1 |
gpu |
1 |
graph |
1 |
graph-ir |
1 |
griffin |
1 |
hardware-aware |
1 |
hqq |
1 |
istft |
1 |
kv |
1 |
language-model |
1 |
latent-diffusion |
1 |
linear |
1 |
llm |
1 |
lookahead-decoding |
1 |
low-bit |
1 |
low-latency |
1 |
low-rank |
1 |
low-token-rate |
1 |
mamba |
1 |
mamba2 |
1 |
media-hash |
1 |
medusa |
1 |
mimi |
1 |
mixtral |
1 |
mla |
1 |
model-config |
1 |
model-detection |
1 |
model-selection |
1 |
mps |
1 |
mqa |
1 |
mrope |
1 |
mtp |
1 |
mx.compile |
1 |
mxfp4 |
1 |
offload |
1 |
onnx |
1 |
packaging |
1 |
padding-mask |
1 |
paged-cache |
1 |
parakeet |
1 |
parameter-efficient |
1 |
peak-memory |
1 |
pipeline |
1 |
pipeline-parallelism |
1 |
projector |
1 |
prompt-cache |
1 |
prompt-lookup |
1 |
publication |
1 |
qwen |
1 |
qwen3-tts |
1 |
rag |
1 |
rectified-flow |
1 |
redrafter |
1 |
regression-tests |
1 |
resampler |
1 |
retrieval |
1 |
retrieval-attention |
1 |
rnnt |
1 |
rotating-cache |
1 |
rvq |
1 |
rwkv |
1 |
security |
1 |
segmented-mm |
1 |
self-speculative |
1 |
semantic |
1 |
semantic-acoustic |
1 |
semantic-tokens |
1 |
serialization |
1 |
sharding |
1 |
side-channel |
1 |
sliding-window |
1 |
smoke-test |
1 |
spatial-warp |
1 |
speech-generation |
1 |
stable-diffusion-3 |
1 |
state-space |
1 |
streaming-speech |
1 |
streamingllm |
1 |
streams |
1 |
sts |
1 |
t5 |
1 |
tacotron2 |
1 |
token-recycling |
1 |
tokenizer |
1 |
toolkit |
1 |
translation |
1 |
valle |
1 |
variable-rate |
1 |
video-diffusion |
1 |
virtual-memory |
1 |
vision-transformer |
1 |
vits |
1 |
voxtral |
1 |
wavernn |
1 |
Records are sorted by stable source ID for review. The canonical registry remains in its original author-curated order.
| ID | Depth | Kind | Evidence class | Support scope | Claim types | Owner | Topics | Snapshot | Source | Claim boundary / note |
|---|---|---|---|---|---|---|---|---|---|---|
agent-skills-repo |
screened |
repository |
not-classified |
not-classified |
— | Anthropic | agent-skills, examples |
57546260929473d4e0d1c1bb75297be2fdfa1949 |
Anthropic Agent Skills repository | — |
agent-skills-spec |
synthesized |
official-doc |
not-classified |
not-classified |
— | Agent Skills | agent-skills, format, validation |
2026-06-24 |
Agent Skills specification | — |
apple-fastvlm-blog |
synthesized |
technical-blog |
not-classified |
not-classified |
— | Apple Machine Learning Research | fastvlm, vision-language, visual-token-reduction |
2026-06-24 |
FastVLM Apple Machine Learning Research blog | Engineering context for FastVLM; do not treat as generic MLX port support. |
apple-metal-guide |
indexed |
official-doc |
not-classified |
not-classified |
— | Apple | compute, gpu, metal |
— |
Metal compute documentation | — |
apple-mlx-m5-blog |
screened |
technical-blog |
not-classified |
not-classified |
— | Apple Machine Learning Research | apple-silicon, llm, mlx, runtime |
2026-06-24 |
Exploring LLMs with MLX and the Neural Accelerators in the M5 GPU | Practitioner context for MLX on Apple Silicon; not a validation substitute. |
apple-mpsgraph |
indexed |
official-doc |
not-classified |
not-classified |
— | Apple | graph, metal, operators |
— |
Metal Performance Shaders Graph documentation | — |
apple-wwdc25-mlx |
screened |
technical-blog |
not-classified |
not-classified |
— | Apple Developer | apple-silicon, compile, lazy-evaluation, mlx |
2026-06-24 |
Get started with MLX for Apple silicon | Developer-session context for MLX concepts; cite official MLX docs for API specifics. |
codex-skill-loader |
synthesized |
source-code |
not-classified |
not-classified |
— | OpenAI | agent-skills, codex |
31e428a1ef3396a0e962687e61f11c224b9f4dbc |
Codex skill loader implementation | — |
dc-dc-dc-mlx-rwkv-repo |
screened |
repository |
not-classified |
not-classified |
— | dc-dc-dc | mlx, recurrent, rwkv, safetensors |
45d9f309afa3 |
mlx-rwkv WIP recurrent implementation | Screened as top-1000 contributor implementation context; held because README marks it WIP and local code still relies on NumPy state shifting. |
dgrauet-ltx2-block-streaming-source |
synthesized |
source-code |
primary_source_code |
third_party_pinned |
memory, mlx_implementation |
dgrauet | block-streaming, diffusion, mlx, peak-memory, safetensors |
3f5897a9582b5c14379c8f88216ae0fd6e55741d |
ltx-2-mlx BlockStreamer implementation | Pinned implementation evidence for memory-mapped repeated-block weight streaming. |
dgrauet-ltx2-block-streaming-tests |
synthesized |
source-code |
not-classified |
not-classified |
— | dgrauet | block-streaming, eviction, mlx, parity, tests |
3f5897a9582b5c14379c8f88216ae0fd6e55741d |
ltx-2-mlx block streaming tests | Pinned correctness evidence for block discovery, weight binding, streamed/eager parity, and eviction/reload behavior. |
dgrauet-ltx2-regression-report |
synthesized |
source-code |
not-classified |
not-classified |
— | dgrauet | block-streaming, mlx, regression-tests, video-diffusion |
3f5897a9582b5c14379c8f88216ae0fd6e55741d |
ltx-2-mlx v0.11.0 regression report | Pinned report claiming byte-identical low-RAM/eager output for a q8 LTX-2 path; source-reported, not locally reproduced here. |
dropbox-hqq-repo |
screened |
repository |
not-classified |
not-classified |
— | Dropbox | hqq, weight-quantization |
2026-06-24 |
Official HQQ implementation | Primary HQQ implementation/source; no MLX support claim by itself. |
fl33twood-mlx-conformer-repo |
screened |
repository |
not-classified |
not-classified |
— | FL33TW00D | asr, conformer, mlx, padding-mask |
88809a360184 |
mlx-conformer Torch-to-MLX port scaffold | Screened as top-1000 contributor implementation context; reinforces existing ASR/Conformer parity needs but lacks a packaged test suite. |
github-mlx-top100-contributors |
screened |
repository |
not-classified |
not-classified |
— | GitHub / ml-explore | contributors, mlx, source-selection |
2026-06-27 |
GitHub API top contributors for ml-explore/mlx | Used to select the top-100 contributor source set for contributor implementation review. |
github-mlx-top1000-contributors |
screened |
repository |
not-classified |
not-classified |
— | GitHub / ml-explore | contributors, mlx, source-selection |
2026-06-27 |
GitHub API top contributors for ml-explore/mlx pages 1-10 | Used to request the top-1000 contributor source set; GitHub returned 256 contributors across the paginated API. Future refreshes should use scripts/collect_contributors.py so Link headers and receipt metadata are preserved. |
goekdeniz-mixtral-from-scratch-repo |
screened |
repository |
not-classified |
not-classified |
— | Goekdeniz-Guelmez | lora, mixtral, mlx, moe |
f08477a06ed9 |
Mixtral from scratch in MLX | Screened as top-1000 contributor implementation context; reinforces MoE loop-oracle and training gates but was not promoted as new guidance. |
hf-mlx-model-search-2026-06-27 |
screened |
repository |
not-classified |
not-classified |
— | Hugging Face | coverage, hugging-face, mlx, source-selection |
2026-06-27 |
Hugging Face MLX model search snapshot | Used for demand and artifact-shape sampling only; downloads and tags are not implementation evidence. |
hf-moss-tts-model-card |
synthesized |
repository |
not-classified |
not-classified |
— | mlx-community | model-card, moss-tts, streaming, tts |
2026-06-24 |
MOSS-TTS MLX community model card | Model card evidence for a specific converted artifact; not generic MLX-Audio capability proof. |
hf-nemotron-diffusion-mlx-card |
screened |
repository |
not-classified |
not-classified |
— | mlx-community | hugging-face, mlx, model-card, publication |
2026-06-27 |
mlx-community Nemotron-Labs-Diffusion-3B-4bit model card | Mutable HF model-card evidence for publication metadata linting, not benchmark or implementation proof. |
hf-top-models-api-2026-06-29 |
screened |
repository |
not-classified |
not-classified |
— | Hugging Face | coverage, downloads, hugging-face, model-selection |
2026-06-29 |
Hugging Face top downloaded public models API snapshot | Used to choose a top-250 public model coverage target; mutable downloads, tags, and license fields are demand signals, not proof of implementation support. |
huggingface-transformers |
screened |
repository |
not-classified |
not-classified |
— | Hugging Face | architectures, model-config, source-oracle |
— |
Transformers architecture and model repository | — |
j-csc-mlx-ssm-repo |
screened |
repository |
not-classified |
not-classified |
— | j-csc | mamba, mlx, ssm, state-space |
51e2eb3285f7 |
mlx-ssm state-space prototype | Screened as top-1000 contributor implementation context; held because key SSM/Mamba kernels are incomplete. |
katlun-grid-sample-benchmark |
synthesized |
source-code |
not-classified |
not-classified |
— | katlun-lgtm | benchmarking, custom-kernel, grid-sample, mlx |
467385fa2b84864659bf3076eca32c4c4d4ddbec |
mlx-grid-sample benchmark harness | Pinned benchmark harness comparing the custom 3D grid_sample kernel with a pure-MLX gather baseline. |
katlun-grid-sample-source |
synthesized |
source-code |
primary_source_code |
third_party_pinned |
mlx_implementation |
katlun-lgtm | custom-kernel, grid-sample, mlx, spatial-warp, video |
467385fa2b84864659bf3076eca32c4c4d4ddbec |
mlx-grid-sample custom Metal implementation | Pinned implementation evidence for PyTorch-compatible 2D/3D grid_sample kernels using mx.fast.metal_kernel. |
katlun-grid-sample-tests |
synthesized |
source-code |
not-classified |
not-classified |
— | katlun-lgtm | custom-kernel, grid-sample, mlx, parity, tests |
467385fa2b84864659bf3076eca32c4c4d4ddbec |
mlx-grid-sample correctness tests | Pinned correctness evidence for align_corners and zeros/border padding parity against NumPy and optional Torch. |
llava-onevision-blog |
synthesized |
technical-blog |
not-classified |
not-classified |
— | LLaVA team | llava, video, vision-language |
2026-06-24 |
LLaVA-OneVision release blog | Context for multi-image/video token budgets and evaluation needs. |
lucasnewman-f5-tts-mlx-repo |
screened |
repository |
not-classified |
not-classified |
— | lucasnewman | duration, flow-matching, mlx, tts, vocoder |
85d015f62b7c |
F5-TTS MLX implementation | Screened as top-100 contributor implementation context; reinforces existing flow-TTS gates. |
lucasnewman-vocos-mlx-repo |
screened |
repository |
not-classified |
not-classified |
— | lucasnewman | audio-codec, encodec, mlx, vocoder |
18c46027a2fb |
Vocos MLX implementation | Screened as top-100 contributor implementation context; reinforces existing vocoder/audio-codec gates. |
mlx-audio-codecs |
synthesized |
source-code |
not-classified |
not-classified |
— | Blaizzy | audio-codec, tokenizer, vocoder |
412cf7cd381c2a3f6a8189af04a95af24cb415b6 |
MLX-Audio codec implementations | — |
mlx-audio-convert |
synthesized |
source-code |
not-classified |
not-classified |
— | Blaizzy | audio, conversion, model-detection, quantization |
412cf7cd381c2a3f6a8189af04a95af24cb415b6 |
MLX-Audio conversion implementation | — |
mlx-audio-docs |
synthesized |
official-doc |
not-classified |
not-classified |
— | Blaizzy | audio, mlx-audio, stt, tts |
2026-06-24 |
MLX-Audio documentation | Landing documentation for MLX-Audio model and API surfaces. |
mlx-audio-models |
synthesized |
source-code |
not-classified |
not-classified |
— | Blaizzy | audio, sts, stt, tts |
412cf7cd381c2a3f6a8189af04a95af24cb415b6 |
MLX-Audio model implementations | — |
mlx-audio-moss-tts-docs |
synthesized |
official-doc |
not-classified |
not-classified |
— | Blaizzy | mlx-audio, moss-tts, streaming, tts |
2026-06-24 |
MLX-Audio MOSS-TTS model guide | Model-specific streaming/codebook context; do not generalize schedule to arbitrary codecs. |
mlx-audio-quantization-guide |
synthesized |
official-doc |
not-classified |
not-classified |
— | Blaizzy | mlx-audio, quantization, stt, tts |
2026-06-24 |
MLX-Audio quantization guide | Source for MLX-Audio bit-width, mode, and quality guidance; still requires model-specific listening/WER gates. |
mlx-audio-qwen3-tts-docs |
synthesized |
source-code |
primary_source_code |
third_party_pinned |
audio_quality, mlx_implementation, performance |
Blaizzy | batching, mlx-audio, qwen3-tts, streaming, tts |
a7ef98604cfd752e9e5c9011bcee8ec8c67228be |
MLX-Audio Qwen3-TTS model guide | Source for Qwen3-TTS streaming and source-reported short-prompt batch throughput table. |
mlx-audio-release-044 |
synthesized |
release |
release_note |
third_party_pinned |
mlx_implementation |
Blaizzy | audio, cache, codec, streaming |
a7ef98604cfd752e9e5c9011bcee8ec8c67228be |
MLX-Audio v0.4.4 release | Pinned release identity for the exact MLX-Audio 0.4.4 software constraint; the performance table is cited separately from the versioned model guide. |
mlx-audio-repo |
synthesized |
repository |
not-classified |
not-classified |
— | Blaizzy | audio, codec, mlx, stt, tts |
412cf7cd381c2a3f6a8189af04a95af24cb415b6 |
MLX-Audio repository | — |
mlx-audio-streaming-guide |
synthesized |
official-doc |
not-classified |
not-classified |
— | Blaizzy | mlx-audio, streaming, stt, tts |
2026-06-24 |
MLX-Audio streaming audio guide | Documents CLI, Python, API, and WebSocket streaming surfaces plus chunk-size trade-offs. |
mlx-audio-swift |
indexed |
repository |
not-classified |
not-classified |
— | Blaizzy | audio, mlx, swift |
— |
MLX-Audio Swift repository | — |
mlx-audio-voxtral-realtime-docs |
synthesized |
official-doc |
not-classified |
not-classified |
— | Blaizzy | mlx-audio, streaming, stt, voxtral |
2026-06-24 |
MLX-Audio Voxtral Realtime STT guide | Model-specific realtime STT API context. |
mlx-dev-skill |
synthesized |
repository |
not-classified |
not-classified |
— | ettrickshepherd | agent-skills, development, mlx |
99e8bd62dd4f0e1649ae9647a1f7ee65b65d02fe |
General MLX development Agent Skill | — |
mlx-doc-async-eval |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | async-eval, mlx, runtime |
2026-06-24 |
MLX async_eval API | Documents asynchronous evaluation API; use with explicit synchronization and determinism gates. |
mlx-doc-block-masked-mm |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | block-sparse, kernels, mlx |
2026-06-24 |
MLX block-masked matmul API | — |
mlx-doc-compile |
synthesized |
official-doc |
official_api_doc |
official_mlx |
api_support |
Apple MLX | compile, fusion, mlx |
2026-06-24 |
MLX compilation documentation | — |
mlx-doc-core-quantize |
synthesized |
official-doc |
official_api_doc |
official_mlx |
api_support |
Apple MLX | low-bit-formats, mlx, quantization |
2026-06-24 |
MLX core quantize API | — |
mlx-doc-cpp-ops |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | cpp, kernels, mlx, segmented-mm |
2026-06-24 |
MLX C++ operations reference | Used only to prove the segmented_mm primitive exists in documented MLX C++; Python and MoE usage still require target validation. |
mlx-doc-custom-extensions |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | custom-kernels, metal, mlx |
2026-06-24 |
MLX custom extensions documentation | — |
mlx-doc-custom-function |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | autograd, custom-function, mlx |
2026-06-24 |
MLX custom_function API | API reference only; custom functions are not a porting shortcut. |
mlx-doc-export |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | export, mlx, packaging, smoke-test |
2026-06-27 |
MLX exporting functions documentation | Official API evidence for optional .mlxfn release smoke tests with shape/dtype caveats. |
mlx-doc-fast-layer-norm |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | kernels, mlx, normalization |
2026-06-24 |
MLX fast LayerNorm API | — |
mlx-doc-fast-metal-kernel |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | custom-kernels, metal, mlx |
2026-06-24 |
MLX fast Metal kernel API | Custom kernels require a reference fallback, shape/dtype coverage, and profile proof. |
mlx-doc-fast-rms-norm |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | kernels, mlx, normalization |
2026-06-24 |
MLX fast RMSNorm API | — |
mlx-doc-fast-rope |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | kernels, mlx, rope |
2026-06-24 |
MLX fast RoPE API | — |
mlx-doc-fast-sdpa |
synthesized |
official-doc |
official_api_doc |
official_mlx |
api_support |
Apple MLX | attention, kernels, mlx |
2026-06-24 |
MLX fast scaled dot product attention API | — |
mlx-doc-gather-mm |
synthesized |
official-doc |
official_api_doc |
official_mlx |
api_support |
Apple MLX | gather, kernels, mlx, moe |
2026-06-24 |
MLX gather matmul API | — |
mlx-doc-gather-qmm |
synthesized |
official-doc |
official_api_doc |
official_mlx |
api_support |
Apple MLX | gather, kernels, mlx, moe, quantization |
2026-06-24 |
MLX gather quantized matmul API | — |
mlx-doc-get-peak-memory |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | benchmarking, memory, mlx |
2026-06-27 |
MLX get_peak_memory API | Official API evidence for peak-memory measurement in benchmark and fit gates. |
mlx-doc-hadamard-transform |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | hadamard, kernels, mlx, quantization |
2026-06-24 |
MLX Hadamard transform API | — |
mlx-doc-lazy |
synthesized |
official-doc |
official_api_doc |
official_mlx |
api_support |
Apple MLX | lazy-evaluation, mlx |
2026-06-24 |
MLX lazy evaluation documentation | — |
mlx-doc-nn-quantize |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | linear, mlx, quantization |
2026-06-24 |
MLX neural network quantize API | — |
mlx-doc-quantized-matmul |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | kernels, mlx, quantization |
2026-06-24 |
MLX quantized matmul API | — |
mlx-doc-set-memory-limit |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | benchmarking, memory, mlx |
2026-06-27 |
MLX set_memory_limit API | Official API evidence for graph-evaluation memory-limit experiments; not a quality or latency proof. |
mlx-doc-streams |
synthesized |
official-doc |
not-classified |
not-classified |
— | Apple MLX | mlx, runtime, streams |
2026-06-24 |
MLX using streams documentation | Runtime guide for stream overlap; not a speedup claim without local timeline evidence. |
mlx-docs |
synthesized |
official-doc |
official_api_doc |
official_mlx |
api_support |
Apple MLX | api, mlx, runtime |
2026-06-24 |
MLX documentation | — |
mlx-examples-repo |
screened |
repository |
not-classified |
not-classified |
— | Apple MLX | diffusion, encodec, examples, whisper |
796f5b53cab69a3d48a44233ce21aae889e94a08 |
MLX examples repository | — |
mlx-lm-cache |
synthesized |
source-code |
primary_source_code |
official_mlx_project |
mlx_implementation |
Apple MLX | kv-cache, quantization, rotating-cache |
2c008fd0252b2c569227d12568356ab88ab0560a |
MLX-LM cache implementations | — |
mlx-lm-convert |
synthesized |
source-code |
primary_source_code |
official_mlx_project |
mlx_implementation |
Apple MLX | conversion, quantization, weights |
2c008fd0252b2c569227d12568356ab88ab0560a |
MLX-LM model conversion and quantization | — |
mlx-lm-issue-956 |
synthesized |
issue-report |
not-classified |
not-classified |
— | Apple MLX | fusion, mlx-lm, moe, quantization |
2026-06-24 |
MLX-LM MoE gate/up fusion proposal | Open proposal with reported speedups; keep experimental until merged and reproduced locally. |
mlx-lm-models |
synthesized |
source-code |
not-classified |
not-classified |
— | Apple MLX | architectures, language-models |
2c008fd0252b2c569227d12568356ab88ab0560a |
MLX-LM architecture implementations | — |
mlx-lm-release-0312 |
synthesized |
release |
not-classified |
not-classified |
— | Apple MLX | correctness-fix, mlx-lm, prompt-cache, speculative-decoding |
v0.31.2 |
MLX-LM v0.31.2 release | Official release evidence that speculative decoding output corruption and prompt-cache behavior changed after the stored MLX-LM 0.31.1 local receipts. |
mlx-lm-release-0313 |
screened |
release |
not-classified |
not-classified |
— | Apple MLX | cache, mlx-lm, threading |
ed1fca4 |
MLX-LM v0.31.3 release | — |
mlx-lm-repo |
synthesized |
repository |
not-classified |
not-classified |
— | Apple MLX | generation, language-models, mlx, training |
2c008fd0252b2c569227d12568356ab88ab0560a |
MLX-LM repository | — |
mlx-lm-speculative |
synthesized |
source-code |
primary_source_code |
official_mlx_project |
mlx_implementation |
Apple MLX | generation, speculative-decoding |
2c008fd0252b2c569227d12568356ab88ab0560a |
MLX-LM speculative generation implementation | — |
mlx-lm-switch-layers |
synthesized |
source-code |
primary_source_code |
official_mlx_project |
mlx_implementation |
Apple MLX | gather-mm, gather-qmm, mlx-lm, moe |
ed1fca4cef15a824c5f1702c80f70b4cffc8e4dd |
MLX-LM Switch/MoE layers implementation | Implementation evidence for SwitchLinear/SwitchGLU-style expert dispatch at the pinned MLX-LM v0.31.3 commit; check the installed revision before use. |
mlx-metal-examples |
screened |
source-code |
not-classified |
not-classified |
— | Apple MLX | custom-kernels, metal |
— |
MLX Metal kernel examples and tests | — |
mlx-release-0312 |
synthesized |
release |
not-classified |
not-classified |
— | Apple MLX | mlx, quantized-matmul, release, threading |
68cf2fd |
MLX v0.31.2 release | — |
mlx-release-0320 |
synthesized |
release |
release_note |
official_mlx |
api_support, mlx_implementation, risk_or_negative |
Apple MLX | correctness, metal, mlx, quantized-matmul, release |
v0.32.0 |
MLX v0.32.0 release | Official primary release evidence for MLX 0.32.0. The release records API additions and correctness fixes, including quantized-matmul, gather, SDPA, compile, Metal, and shape-overflow paths; it does not prove a target-model speedup. Reviewed 2026-07-09. |
mlx-repo |
synthesized |
repository |
not-classified |
not-classified |
— | Apple MLX | cuda, metal, mlx, runtime |
96296e9c3075a2389bc5c0f078bf01b5aa377cd9 |
MLX array framework repository | — |
mlx-swift |
screened |
repository |
not-classified |
not-classified |
— | Apple MLX | mlx, swift |
— |
MLX Swift repository | — |
mlx-vlm-apc-source |
synthesized |
source-code |
not-classified |
not-classified |
— | Blaizzy | automatic-prefix-cache, cache, mlx-vlm, multimodal |
6a8cdff6a1f53f46a15d4adb997c3b2d5f621263 |
MLX-VLM automatic prefix cache implementation | Block hash chain supports multimodal extra hashes and tenant scoping in this pinned implementation. |
mlx-vlm-ar-cache-guards |
synthesized |
source-code |
not-classified |
not-classified |
— | Blaizzy | cache-safety, media-hash, mlx-vlm, prefix-cache |
6a8cdff6a1f53f46a15d4adb997c3b2d5f621263 |
MLX-VLM generation cache media guards | Implementation evidence for media payload hashes and prefix safety checks. |
mlx-vlm-readme |
synthesized |
source-code |
primary_source_code |
third_party_pinned |
mlx_implementation, serving_semantics |
Blaizzy | continuous-batching, kv-quantization, prefix-cache, speculative-decoding |
6a8cdff6a1f53f46a15d4adb997c3b2d5f621263 |
MLX-VLM serving, caching, and speculative decoding documentation | — |
mlx-vlm-repo |
synthesized |
repository |
not-classified |
not-classified |
— | Blaizzy | mlx, multimodal, serving |
6a8cdff6a1f53f46a15d4adb997c3b2d5f621263 |
MLX-VLM repository | — |
mlx-vlm-vision-cache-source |
synthesized |
source-code |
primary_source_code |
third_party_pinned |
mlx_implementation, serving_semantics |
Blaizzy | cache, mlx-vlm, multimodal, vision-cache |
6a8cdff6a1f53f46a15d4adb997c3b2d5f621263 |
MLX-VLM vision feature cache implementation | Projected-feature cache implementation with bounded LRU; cache key validation remains mandatory. |
mzbac-mlx-moe-repo |
screened |
repository |
not-classified |
not-classified |
— | mzbac | lora, mlx, moe, quantization |
11088f36eaea |
mlx-moe dense-to-MoE assembly scripts | Screened as top-1000 contributor implementation context; held as synthetic MoE assembly rather than source-preserving porting guidance. |
mzbac-mlx-sharding-repo |
screened |
repository |
not-classified |
not-classified |
— | mzbac | distributed-inference, mlx, pipeline-parallelism, sharding |
8b01c33cae6c |
mlx_sharding distributed inference demo | Screened as top-100 contributor implementation context; held as research until parity, scheduler, failure, and security gates exist. |
nvidia-nvfp4-blog |
screened |
technical-blog |
not-classified |
not-classified |
— | NVIDIA | low-bit-formats, nvfp4, quantization |
2026-06-24 |
Introducing NVFP4 for Efficient and Accurate Low-Precision Inference | Format background for NVFP4; hardware claims are NVIDIA-specific and not MLX proof. |
openclaw-mlx-audio |
screened |
repository |
not-classified |
not-classified |
— | cosformula | agent-skills, audio, tts |
— |
OpenClaw MLX-Audio integration | — |
orca-osdi22 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | continuous-batching, serving |
OSDI 2022 |
Orca: A Distributed Serving System for Transformer-Based Generative Models | Primary USENIX OSDI 2022 source; arXiv:2206.02658 is unrelated. |
paper-1609-03499 |
indexed |
paper |
not-classified |
not-classified |
— | authors | audio, autoregressive, vocoder |
1609.03499 |
WaveNet: A Generative Model for Raw Audio | — |
paper-1703-10135 |
indexed |
paper |
not-classified |
not-classified |
— | authors | tts |
1703.10135 |
Tacotron: Towards End-to-End Speech Synthesis | — |
paper-1706-03762 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | attention, transformer |
1706.03762 |
Attention Is All You Need | — |
paper-1712-05884 |
indexed |
paper |
not-classified |
not-classified |
— | authors | tacotron2, tts |
1712.05884 |
Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions | — |
paper-1802-08435 |
indexed |
paper |
not-classified |
not-classified |
— | authors | vocoder, wavernn |
1802.08435 |
Efficient Neural Audio Synthesis | — |
paper-1810-04805 |
screened |
paper |
not-classified |
not-classified |
— | authors | encoder-transformer |
1810.04805 |
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding | — |
paper-1810-11846 |
indexed |
paper |
not-classified |
not-classified |
— | authors | on-device, vocoder |
1810.11846 |
LPCNet: Improving Neural Speech Synthesis Through Linear Prediction | — |
paper-1811-00002 |
indexed |
paper |
not-classified |
not-classified |
— | authors | flow, vocoder |
1811.00002 |
WaveGlow: A Flow-based Generative Network for Speech Synthesis | — |
paper-1905-09263 |
indexed |
paper |
not-classified |
not-classified |
— | authors | non-autoregressive, tts |
1905.09263 |
FastSpeech: Fast, Robust and Controllable Text to Speech | — |
paper-1910-06711 |
indexed |
paper |
not-classified |
not-classified |
— | authors | gan, vocoder |
1910.06711 |
MelGAN: Generative Adversarial Networks for Conditional Waveform Synthesis | — |
paper-1910-10683 |
screened |
paper |
not-classified |
not-classified |
— | authors | encoder-decoder, t5 |
1910.10683 |
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer | — |
paper-1910-11480 |
indexed |
paper |
not-classified |
not-classified |
— | authors | gan, vocoder |
1910.11480 |
Parallel WaveGAN: A Fast Waveform Generation Model Based on Generative Adversarial Networks | — |
paper-1910-13461 |
screened |
paper |
not-classified |
not-classified |
— | authors | bart, encoder-decoder |
1910.13461 |
BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension | — |
paper-1911-02150 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | attention, mqa |
1911.02150 |
Fast Transformer Decoding: One Write-Head is All You Need | — |
paper-2002-02562 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | asr, rnnt, streaming |
2002.02562 |
Transformer Transducer: A Streamable Speech Recognition Model with Transformer Encoders and RNN-T Loss | — |
paper-2005-08100 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | asr, conformer |
2005.08100 |
Conformer: Convolution-augmented Transformer for Speech Recognition | — |
paper-2005-11129 |
indexed |
paper |
not-classified |
not-classified |
— | authors | flow, tts |
2005.11129 |
Glow-TTS: A Generative Flow for Text-to-Speech via Monotonic Alignment Search | — |
paper-2006-04558 |
indexed |
paper |
not-classified |
not-classified |
— | authors | non-autoregressive, tts |
2006.04558 |
FastSpeech 2: Fast and High-Quality End-to-End Text to Speech | — |
paper-2006-11477 |
screened |
paper |
not-classified |
not-classified |
— | authors | asr, speech-encoder |
2006.11477 |
wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations | — |
paper-2009-00713 |
indexed |
paper |
not-classified |
not-classified |
— | authors | diffusion, vocoder |
2009.00713 |
WaveGrad: Estimating Gradients for Waveform Generation | — |
paper-2009-09761 |
indexed |
paper |
not-classified |
not-classified |
— | authors | audio, diffusion, vocoder |
2009.09761 |
DiffWave: A Versatile Diffusion Model for Audio Synthesis | — |
paper-2010-05646 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | gan, vocoder |
2010.05646 |
HiFi-GAN: Generative Adversarial Networks for Efficient and High Fidelity Speech Synthesis | — |
paper-2010-11929 |
screened |
paper |
not-classified |
not-classified |
— | authors | vision-transformer |
2010.11929 |
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale | — |
paper-2105-06337 |
indexed |
paper |
not-classified |
not-classified |
— | authors | diffusion, tts |
2105.06337 |
Grad-TTS: A Diffusion Probabilistic Model for Text-to-Speech | — |
paper-2106-06103 |
screened |
paper |
not-classified |
not-classified |
— | authors | tts, vits |
2106.06103 |
Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech | — |
paper-2106-07447 |
screened |
paper |
not-classified |
not-classified |
— | authors | asr, speech-encoder |
2106.07447 |
HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units | — |
paper-2106-07889 |
indexed |
paper |
not-classified |
not-classified |
— | authors | vocoder |
2106.07889 |
UnivNet: A Neural Vocoder with Multi-Resolution Spectrogram Discriminators for High-Fidelity Waveform Generation | — |
paper-2106-09685 |
screened |
paper |
not-classified |
not-classified |
— | authors | lora, parameter-efficient, training |
2106.09685 |
LoRA: Low-Rank Adaptation of Large Language Models | — |
paper-2107-03312 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-codec, rvq, streaming |
2107.03312 |
SoundStream: An End-to-End Neural Audio Codec | — |
paper-2110-13900 |
screened |
paper |
not-classified |
not-classified |
— | authors | asr, speech-encoder |
2110.13900 |
WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing | — |
paper-2112-10752 |
screened |
paper |
not-classified |
not-classified |
— | authors | latent-diffusion |
2112.10752 |
High-Resolution Image Synthesis with Latent Diffusion Models | — |
paper-2203-02395 |
screened |
paper |
not-classified |
not-classified |
— | authors | istft, vocoder |
2203.02395 |
iSTFTNet: Fast and Lightweight Mel-Spectrogram Vocoder Incorporating Inverse Short-Time Fourier Transform | — |
paper-2204-14198 |
screened |
paper |
not-classified |
not-classified |
— | authors | resampler, vision-language |
2204.14198 |
Flamingo: a Visual Language Model for Few-Shot Learning | — |
paper-2205-04421 |
indexed |
paper |
not-classified |
not-classified |
— | authors | tts |
2205.04421 |
NaturalSpeech: End-to-End Text to Speech Synthesis with Human-Level Quality | — |
paper-2205-14135 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | attention, kernel |
2205.14135 |
FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness | — |
paper-2206-01861 |
screened |
paper |
not-classified |
not-classified |
— | authors | quantization |
2206.01861 |
ZeroQuant: Efficient and Affordable Post-Training Quantization for Large-Scale Transformers | — |
paper-2206-04658 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | gan, vocoder |
2206.04658 |
BigVGAN: A Universal Neural Vocoder with Large-Scale Training | — |
paper-2206-08317 |
screened |
paper |
not-classified |
not-classified |
— | authors | asr, non-autoregressive |
2206.08317 |
Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition | — |
paper-2208-07339 |
screened |
paper |
not-classified |
not-classified |
— | authors | outliers, quantization |
2208.07339 |
LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale | — |
paper-2209-03003 |
screened |
paper |
not-classified |
not-classified |
— | authors | rectified-flow |
2209.03003 |
Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow | — |
paper-2209-03143 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-lm, codec-tokens |
2209.03143 |
AudioLM: a Language Modeling Approach to Audio Generation | — |
paper-2209-15352 |
screened |
paper |
not-classified |
not-classified |
— | authors | audio-generation, codec-tokens |
2209.15352 |
AudioGen: Textually Guided Audio Generation | — |
paper-2210-02747 |
screened |
paper |
not-classified |
not-classified |
— | authors | flow-matching |
2210.02747 |
Flow Matching for Generative Modeling | — |
paper-2210-13438 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-codec, encodec |
2210.13438 |
High Fidelity Neural Audio Compression | — |
paper-2210-17323 |
screened |
paper |
not-classified |
not-classified |
— | authors | weight-quantization |
2210.17323 |
GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers | — |
paper-2211-10438 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | activation-quantization |
2211.10438 |
SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models | — |
paper-2211-15841 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | block-sparse, moe, training |
2211.15841 |
MegaBlocks: Efficient Sparse Training with Mixture-of-Experts | Research context for block-sparse MoE; not evidence of MLX inference support. |
paper-2211-17192 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | speculative-decoding |
2211.17192 |
Fast Inference from Transformers via Speculative Decoding | — |
paper-2212-04356 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | asr, whisper |
2212.04356 |
Robust Speech Recognition via Large-Scale Weak Supervision | — |
paper-2212-09748 |
screened |
paper |
not-classified |
not-classified |
— | authors | diffusion, dit |
2212.09748 |
Scalable Diffusion Models with Transformers | — |
paper-2301-02111 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-codec, tts, valle |
2301.02111 |
Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers | — |
paper-2301-12597 |
screened |
paper |
not-classified |
not-classified |
— | authors | projector, vision-language |
2301.12597 |
BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models | — |
paper-2302-01318 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | speculative-decoding |
2302.01318 |
Accelerating Large Language Model Decoding with Speculative Sampling | — |
paper-2302-10866 |
indexed |
paper |
not-classified |
not-classified |
— | authors | convolution, long-context |
2302.10866 |
Hyena Hierarchy: Towards Larger Convolutional Language Models | — |
paper-2302-13971 |
screened |
paper |
not-classified |
not-classified |
— | authors | decoder-transformer |
2302.13971 |
LLaMA: Open and Efficient Foundation Language Models | — |
paper-2304-08485 |
screened |
paper |
not-classified |
not-classified |
— | authors | llava, vision-language |
2304.08485 |
Visual Instruction Tuning | — |
paper-2304-09116 |
screened |
paper |
not-classified |
not-classified |
— | authors | diffusion, tts |
2304.09116 |
NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers | — |
paper-2305-09636 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-generation, parallel-codebooks |
2305.09636 |
SoundStorm: Efficient Parallel Audio Generation | — |
paper-2305-09781 |
screened |
paper |
not-classified |
not-classified |
— | authors | serving, speculative-decoding |
2305.09781 |
SpecInfer: Accelerating Generative Large Language Model Serving with Tree-based Speculative Inference and Verification | — |
paper-2305-13048 |
screened |
paper |
not-classified |
not-classified |
— | authors | language-model, recurrent |
2305.13048 |
RWKV: Reinventing RNNs for the Transformer Era | — |
paper-2305-13245 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | attention, gqa |
2305.13245 |
GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints | — |
paper-2305-14314 |
screened |
paper |
not-classified |
not-classified |
— | authors | lora, quantization, training |
2305.14314 |
QLoRA: Efficient Finetuning of Quantized LLMs | — |
paper-2305-17118 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-pruning |
2305.17118 |
Scissorhands: Exploiting the Persistence of Importance Hypothesis for LLM KV Cache Compression | — |
paper-2306-00814 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | fourier, vocoder |
2306.00814 |
Vocos: Closing the Gap between Time-Domain and Fourier-Based Neural Vocoders for High-Quality Audio Synthesis | — |
paper-2306-00978 |
screened |
paper |
not-classified |
not-classified |
— | authors | weight-quantization |
2306.00978 |
AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration | — |
paper-2306-03078 |
screened |
paper |
not-classified |
not-classified |
— | authors | outliers, weight-quantization |
2306.03078 |
SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression | — |
paper-2306-05284 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-lm, delay-pattern, musicgen |
2306.05284 |
Simple and Controllable Music Generation | — |
paper-2306-06546 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-codec, dac |
2306.06546 |
High-Fidelity Audio Compression with Improved RVQGAN | — |
paper-2306-07629 |
screened |
paper |
not-classified |
not-classified |
— | authors | weight-quantization |
2306.07629 |
SqueezeLLM: Dense-and-Sparse Quantization | — |
paper-2306-14048 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-pruning, long-context |
2306.14048 |
H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models | — |
paper-2306-15687 |
screened |
paper |
not-classified |
not-classified |
— | authors | flow, speech-generation |
2306.15687 |
Voicebox: Text-Guided Multilingual Universal Speech Generation at Scale | — |
paper-2307-08621 |
indexed |
paper |
not-classified |
not-classified |
— | authors | linear-attention, recurrent |
2307.08621 |
Retentive Network: A Successor to Transformer for Large Language Models | — |
paper-2307-08691 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | attention, kernel |
2307.08691 |
FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning | — |
paper-2307-09288 |
screened |
paper |
not-classified |
not-classified |
— | authors | decoder-transformer |
2307.09288 |
Llama 2: Open Foundation and Fine-Tuned Chat Models | — |
paper-2308-11596 |
screened |
paper |
not-classified |
not-classified |
— | authors | asr, speech-to-speech, translation |
2308.11596 |
SeamlessM4T: Massively Multilingual & Multimodal Machine Translation | — |
paper-2308-13137 |
screened |
paper |
not-classified |
not-classified |
— | authors | quantization |
2308.13137 |
OmniQuant: Omnidirectionally Calibrated Quantization for Large Language Models | — |
paper-2308-16692 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-codec, speech-tokenizer |
2308.16692 |
SpeechTokenizer: Unified Speech Tokenizer for Speech Language Models | — |
paper-2309-03199 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | flow-matching, tts |
2309.03199 |
Matcha-TTS: A Fast TTS Architecture with Conditional Flow Matching | — |
paper-2309-06180 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | paged-attention, serving |
2309.06180 |
Efficient Memory Management for Large Language Model Serving with PagedAttention | — |
paper-2309-07405 |
screened |
paper |
not-classified |
not-classified |
— | authors | audio-codec, toolkit |
2309.07405 |
FunCodec: A Fundamental, Reproducible and Integrable Open-source Toolkit for Neural Speech Codec | — |
paper-2309-08168 |
screened |
paper |
not-classified |
not-classified |
— | authors | self-speculative |
2309.08168 |
Draft & Verify: Lossless Large Language Model Acceleration via Self-Speculative Decoding | — |
paper-2309-15505 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-codec, fsq, quantization |
2309.15505 |
Finite Scalar Quantization: VQ-VAE Made Simple | — |
paper-2309-15531 |
screened |
paper |
not-classified |
not-classified |
— | authors | outliers, weight-quantization |
2309.15531 |
Rethinking Channel Dimensions to Isolate Outliers for Low-bit Weight Quantization of Large Language Models | Not HQQ; this is the AdaDim paper. |
paper-2309-17453 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | attention-sinks, streamingllm |
2309.17453 |
Efficient Streaming Language Models with Attention Sinks | — |
paper-2310-00014 |
screened |
paper |
not-classified |
not-classified |
— | authors | audio-codec, tts |
2310.00014 |
Fewer-token Neural Speech Codec with Time-invariant Codes | — |
paper-2310-01889 |
screened |
paper |
not-classified |
not-classified |
— | authors | distributed-attention, long-context |
2310.01889 |
Ring Attention with Blockwise Transformers for Near-Infinite Context | — |
paper-2310-06825 |
screened |
paper |
not-classified |
not-classified |
— | authors | decoder-transformer, gqa, sliding-window |
2310.06825 |
Mistral 7B | — |
paper-2310-13289 |
screened |
paper |
not-classified |
not-classified |
— | authors | audio-language, multimodal |
2310.13289 |
SALMONN: Towards Generic Hearing Abilities for Large Language Models | — |
paper-2310-16836 |
screened |
paper |
not-classified |
not-classified |
— | authors | fp4, quantization |
2310.16836 |
LLM-FP4: 4-Bit Floating-Point Quantized Transformers | — |
paper-2311-00430 |
screened |
paper |
not-classified |
not-classified |
— | authors | asr, distillation |
2311.00430 |
Distil-Whisper: Robust Knowledge Distillation via Large-Scale Pseudo Labelling | — |
paper-2311-08252 |
screened |
paper |
not-classified |
not-classified |
— | authors | retrieval, speculative-decoding |
2311.08252 |
REST: Retrieval-Based Speculative Decoding | — |
paper-2312-00752 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | recurrent, ssm |
2312.00752 |
Mamba: Linear-Time Sequence Modeling with Selective State Spaces | — |
paper-2312-03788 |
screened |
paper |
not-classified |
not-classified |
— | authors | quantization |
2312.03788 |
SmoothQuant+: Accurate and Efficient 4-bit Post-Training Weight Quantization for LLM | — |
paper-2312-05187 |
screened |
paper |
not-classified |
not-classified |
— | authors | speech-to-speech, streaming |
2312.05187 |
Seamless: Multilingual Expressive and Streaming Speech Translation | — |
paper-2312-17279 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | asr, cache, streaming |
2312.17279 |
Stateful Conformer with Cache-based Inference for Streaming Automatic Speech Recognition | — |
paper-2401-04088 |
screened |
paper |
not-classified |
not-classified |
— | authors | decoder-transformer, moe |
2401.04088 |
Mixtral of Experts | — |
paper-2401-04577 |
screened |
paper |
not-classified |
not-classified |
— | authors | audio-generation, non-autoregressive |
2401.04577 |
MAGNeT: Masked Audio Generation using a Single Non-Autoregressive Transformer | — |
paper-2401-06118 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | quantization |
2401.06118 |
AQLM: Extreme Compression of Large Language Models via Additive Quantization | — |
paper-2401-07851 |
screened |
paper |
not-classified |
not-classified |
— | authors | speculative-decoding, survey |
2401.07851 |
Unlocking Efficiency in Large Language Model Inference: A Comprehensive Survey of Speculative Decoding | — |
paper-2401-10774 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | medusa, speculative-decoding |
2401.10774 |
Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads | — |
paper-2401-15077 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | eagle, speculative-decoding |
2401.15077 |
EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty | — |
paper-2401-18079 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | kv-quantization |
2401.18079 |
KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization | — |
paper-2402-01831 |
screened |
paper |
not-classified |
not-classified |
— | authors | audio-language, multimodal |
2402.01831 |
Audio Flamingo: A Novel Audio Language Model with Few-Shot Learning and Dialogue Abilities | — |
paper-2402-02057 |
screened |
paper |
not-classified |
not-classified |
— | authors | lookahead-decoding |
2402.02057 |
Break the Sequential Dependency of LLM Inference Using Lookahead Decoding | — |
paper-2402-02750 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | kv-quantization |
2402.02750 |
KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache | — |
paper-2402-04396 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | hadamard, quantization |
2402.04396 |
QuIP#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks | — |
paper-2402-05099 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | prefix-sharing, serving |
2402.05099 |
Hydragen: High-Throughput LLM Inference with Shared Prefixes | — |
paper-2402-06082 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-compression |
2402.06082 |
SubGen: Token Generation in Sublinear Time and Memory | — |
paper-2402-11131 |
screened |
paper |
not-classified |
not-classified |
— | authors | multi-token, speculative-decoding |
2402.11131 |
Speculative Streaming: Fast LLM Inference without Auxiliary Models | — |
paper-2402-13236 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-language, codec, survey |
2402.13236 |
Towards Audio Language Modeling: An Overview | — |
paper-2402-13720 |
screened |
paper |
not-classified |
not-classified |
— | authors | speculative-decoding |
2402.13720 |
Ouroboros: Speculative Decoding with Large Model Enhanced Drafting | — |
paper-2402-15220 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | attention, prefix-cache |
2402.15220 |
ChunkAttention: Efficient Self-Attention with Prefix-Aware KV Cache and Two-Phase Partition | — |
paper-2402-17764 |
screened |
paper |
not-classified |
not-classified |
— | authors | bitnet, quantization |
2402.17764 |
The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits | — |
paper-2402-19427 |
screened |
paper |
not-classified |
not-classified |
— | authors | hybrid, recurrent |
2402.19427 |
Griffin: Mixing Gated Linear Recurrences with Local Attention for Efficient Language Models | — |
paper-2403-03206 |
screened |
paper |
not-classified |
not-classified |
— | authors | flow, stable-diffusion-3, transformer |
2403.03206 |
Scaling Rectified Flow Transformers for High-Resolution Image Synthesis | — |
paper-2403-08295 |
screened |
paper |
not-classified |
not-classified |
— | authors | decoder-transformer |
2403.08295 |
Gemma: Open Models Based on Gemini Research and Technology | — |
paper-2403-09919 |
screened |
paper |
not-classified |
not-classified |
— | authors | mlx, redrafter, speculative-decoding |
2403.09919 |
ReDrafter: Faster LLM Inference with Recurrent Drafter | — |
paper-2403-15388 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | llava, vision-language, visual-token-pruning |
2403.15388 |
LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models | Visual-token pruning research; not generic support. |
paper-2403-19887 |
screened |
paper |
not-classified |
not-classified |
— | authors | attention, hybrid, moe, ssm |
2403.19887 |
Jamba: A Hybrid Transformer-Mamba Language Model | — |
paper-2404-00456 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | kv, quantization, rotation |
2404.00456 |
QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs | — |
paper-2404-02781 |
screened |
paper |
not-classified |
not-classified |
— | authors | codec-lm, tts |
2404.02781 |
CLaM-TTS: Improving Neural Codec Language Model for Zero-Shot Text-to-Speech | — |
paper-2404-07839 |
screened |
paper |
not-classified |
not-classified |
— | authors | griffin, recurrent |
2404.07839 |
RecurrentGemma: Moving Past Transformers for Efficient Open Language Models | — |
paper-2404-14219 |
screened |
paper |
not-classified |
not-classified |
— | authors | decoder-transformer, on-device |
2404.14219 |
Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone | — |
paper-2404-14469 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-pruning |
2404.14469 |
SnapKV: LLM Knows What You Are Looking for Before Generation | — |
paper-2404-19124 |
screened |
paper |
not-classified |
not-classified |
— | authors | multi-token, speculative-decoding |
2404.19124 |
Accelerating Production LLMs with Combined Token/Embedding Speculators | Not Prompt Lookup Decoding; use prompt-lookup-decoding-repo for PLD. |
paper-2405-04434 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | attention, mla, moe |
2405.04434 |
DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model | — |
paper-2405-04437 |
screened |
paper |
not-classified |
not-classified |
— | authors | serving, virtual-memory |
2405.04437 |
vAttention: Dynamic Memory Management for Serving LLMs without PagedAttention | — |
paper-2405-04532 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | quantization, serving |
2405.04532 |
QServe: W4A8KV4 Quantization and System Co-design for Efficient LLM Serving | — |
paper-2405-16406 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | quantization, rotation |
2405.16406 |
SpinQuant: LLM Quantization with Learned Rotations | — |
paper-2405-21060 |
screened |
paper |
not-classified |
not-classified |
— | authors | mamba2, ssm |
2405.21060 |
Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality | — |
paper-2406-02069 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-compression |
2406.02069 |
PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling | — |
paper-2406-03482 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-quantization |
2406.03482 |
QJL: 1-Bit Quantized JL Transform for KV Cache Quantization with Zero Overhead | — |
paper-2406-16858 |
screened |
paper |
not-classified |
not-classified |
— | authors | eagle2, speculative-decoding |
2406.16858 |
EAGLE-2: Faster Inference of Language Models with Dynamic Draft Trees | — |
paper-2406-18009 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | flow-matching, tts |
2406.18009 |
E2 TTS: Embarrassingly Easy Fully Non-Autoregressive Zero-Shot TTS | — |
paper-2407-00088 |
screened |
paper |
not-classified |
not-classified |
— | authors | edge, low-bit |
2407.00088 |
T-MAC: CPU Renaissance via Table Lookup for Low-Bit LLM Deployment on Edge | — |
paper-2407-01955 |
screened |
paper |
not-classified |
not-classified |
— | authors | speculative-decoding |
2407.01955 |
S2D: Sorted Speculative Decoding for More Efficient Deployment of Nested Large Language Models | — |
paper-2407-08608 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | attention, kernel |
2407.08608 |
FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-Precision | — |
paper-2407-10671 |
screened |
paper |
not-classified |
not-classified |
— | authors | decoder-transformer, moe |
2407.10671 |
Qwen2 Technical Report | — |
paper-2407-10759 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-language, multimodal |
2407.10759 |
Qwen2-Audio Technical Report | — |
paper-2407-17678 |
screened |
paper |
not-classified |
not-classified |
— | authors | context-sharding, sparse-attention |
2407.17678 |
Efficient LLM Training and Serving with Heterogeneous Context Sharding among Attention Heads | — |
paper-2407-21783 |
screened |
paper |
not-classified |
not-classified |
— | authors | decoder-transformer |
2407.21783 |
The Llama 3 Herd of Models | — |
paper-2408-00118 |
screened |
paper |
not-classified |
not-classified |
— | authors | decoder-transformer |
2408.00118 |
Gemma 2: Improving Open Language Models at a Practical Size | — |
paper-2408-05646 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-compression, low-rank |
2408.05646 |
Eigen Attention: Attention in Low-Rank Space for KV Cache Compression | — |
paper-2408-08554 |
screened |
paper |
not-classified |
not-classified |
— | authors | quantization |
2408.08554 |
ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models | — |
paper-2408-08696 |
screened |
paper |
not-classified |
not-classified |
— | authors | speculative-decoding, token-recycling |
2408.08696 |
Turning Trash into Treasure: Accelerating Inference of Large Language Models with Token Recycling | — |
paper-2408-16532 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-codec, low-token-rate |
2408.16532 |
WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling | — |
paper-2409-00142 |
screened |
paper |
not-classified |
not-classified |
— | authors | speculative-decoding |
2409.00142 |
Dynamic Depth Decoding: Faster Speculative Decoding for LLMs | — |
paper-2409-12191 |
screened |
paper |
not-classified |
not-classified |
— | authors | video, vision-language |
2409.12191 |
Qwen2-VL: Enhancing Vision-Language Model’s Perception of the World at Any Resolution | — |
paper-2410-00037 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | mimi, speech-to-speech, streaming |
2410.00037 |
Moshi: a Speech-Text Foundation Model for Real-Time Dialogue | — |
paper-2410-03804 |
screened |
paper |
not-classified |
not-classified |
— | authors | attention, speculative-decoding |
2410.03804 |
Mixture of Attentions for Speculative Decoding | — |
paper-2410-06885 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | flow-matching, tts |
2410.06885 |
F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching | — |
paper-2410-07505 |
screened |
paper |
not-classified |
not-classified |
— | authors | activation-quantization |
2410.07505 |
CrossQuant: A Post-Training Quantization Method with Smaller Quantization Kernel for Precise Large Language Model Compression | — |
paper-2410-12168 |
screened |
paper |
not-classified |
not-classified |
— | authors | quantization, serving |
2410.12168 |
COMET: Towards Practical W4A4KV4 LLMs Serving | — |
paper-2411-04965 |
screened |
paper |
not-classified |
not-classified |
— | authors | activation-quantization, bitnet |
2411.04965 |
BitNet a4.8: 4-bit Activations for 1-bit LLMs | — |
paper-2411-18803 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-codec, streaming |
2411.18803 |
TS3-Codec: Transformer-Based Simple Streaming Single Codec | — |
paper-2411-19842 |
screened |
paper |
not-classified |
not-classified |
— | authors | audio-codec, fsq |
2411.19842 |
Scaling Transformers for Low-Bitrate High-Quality Speech Coding | — |
paper-2412-03594 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | batching, prefix-sharing |
2412.03594 |
BatchLLM: Optimizing Large Batched LLM Inference with Global Prefix Sharing | — |
paper-2412-04467 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | vision-language, visual-token-pruning |
2412.04467 |
VisionZip: Longer is Better but Not Necessary in Vision Language Models | Visual-token compression research; requires per-model quality and MLX implementation gates. |
paper-2412-10319 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | benchmark, kv-cache, long-context |
2412.10319 |
SCBench: A KV Cache-Centric Analysis of Long-Context Methods | — |
paper-2412-12639 |
screened |
paper |
not-classified |
not-classified |
— | authors | speculative-decoding |
2412.12639 |
Falcon: Faster and Parallel Inference of Large Language Models through Enhanced Semi-Autoregressive Drafting and Custom-Designed Decoding Tree | — |
paper-2412-13303 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | encoder, vision-language, visual-token-reduction |
2412.13303 |
FastVLM: Efficient Vision Encoding for Vision Language Models | Research source for efficient VLM encoders; replacing encoders is a model change, not a serving optimization. |
paper-2412-14363 |
screened |
paper |
not-classified |
not-classified |
— | authors | mixed-precision, quantization |
2412.14363 |
ResQ: Mixed-Precision Quantization of Large Language Models with Low-Rank Residuals | — |
paper-2412-19437 |
screened |
paper |
not-classified |
not-classified |
— | authors | attention, moe, mtp |
2412.19437 |
DeepSeek-V3 Technical Report | — |
paper-2501-01005 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | attention, kernels, serving |
2501.01005 |
FlashInfer: Efficient and Customizable Attention Engine for LLM Inference Serving | — |
paper-2501-16383 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-quantization, rotation |
2501.16383 |
RotateKV: Accurate and Robust 2-Bit KV Cache Quantization via Outlier-Aware Adaptive Rotations | — |
paper-2502-04077 |
screened |
paper |
not-classified |
not-classified |
— | authors | attention, kv-compression |
2502.04077 |
AttentionPredictor: Temporal Pattern Matters for Efficient LLM Inference | — |
paper-2502-04420 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | kv-quantization, mixed-precision |
2502.04420 |
KVTuner: Sensitivity-Aware Layer-Wise Mixed-Precision KV Cache Quantization | — |
paper-2502-13923 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | mrope, qwen, video, vision-language |
2502.13923 |
Qwen2.5-VL Technical Report | Research source for Qwen visual token/grid/M-RoPE semantics; not MLX support by itself. |
paper-2502-15075 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-quantization, mixed-precision |
2502.15075 |
More for Keys, Less for Values: Adaptive KV Cache Quantization | — |
paper-2503-01710 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | codec-lm, tts |
2503.01710 |
Spark-TTS: An Efficient LLM-Based Text-to-Speech Model with Single-Stream Decoupled Speech Tokens | — |
paper-2503-14376 |
screened |
paper |
not-classified |
not-classified |
— | authors | kernel, linear-attention |
2503.14376 |
Tiled Flash Linear Attention: More Efficient Linear RNN and xLSTM Kernels | — |
paper-2503-16257 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-quantization, vlm |
2503.16257 |
Plug-and-Play 1.x-Bit KV Cache Quantization for Video Large Language Models | — |
paper-2503-19950 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-quantization |
2503.19950 |
LogQuant: Log-Distributed 2-Bit Quantization of KV Cache | — |
paper-2503-20215 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | omni, speech, streaming |
2503.20215 |
Qwen2.5-Omni Technical Report | — |
paper-2504-10344 |
screened |
paper |
not-classified |
not-classified |
— | authors | audio-codec, semantic-tokens |
2504.10344 |
ALMTokenizer: A Low-bitrate and Semantic-rich Audio Codec Tokenizer for Audio Language Modeling | — |
paper-2504-12339 |
screened |
paper |
not-classified |
not-classified |
— | authors | multi-token, tts |
2504.12339 |
GOAT-TTS: LLM-based Text-To-Speech Generation Optimized via a Dual-Branch Architecture | — |
paper-2504-19874 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-cache, long-context, mlx-vlm, quantization |
2504.19874 |
TurboQuant: Ultra-Efficient and Accurate KV Cache Quantization for Long Context LLM Inference | Paper evidence for TurboQuant research context; not official MLX core support. |
paper-2505-03745 |
screened |
paper |
not-classified |
not-classified |
— | authors | long-context, quantization |
2505.03745 |
AccLLM: Accelerating Long-Context LLM Inference via Algorithm-Hardware Co-Design | — |
paper-2505-07858 |
screened |
paper |
not-classified |
not-classified |
— | authors | scaling-laws, speculative-decoding |
2505.07858 |
Scaling Laws for Speculative Decoding | — |
paper-2505-15380 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | speculative-decoding, speech |
2505.15380 |
Accelerating Autoregressive Speech Synthesis Inference With Speech Speculative Decoding | — |
paper-2505-19201 |
screened |
paper |
not-classified |
not-classified |
— | authors | multimodal, speculative-decoding |
2505.19201 |
DREAM: Drafting with Refined Target Features and Entropy-Adaptive Cross-Attention Fusion for Multimodal Speculative Decoding | — |
paper-2505-19645 |
screened |
paper |
not-classified |
not-classified |
— | authors | moe, speculative-decoding |
2505.19645 |
MoESD: Unveil Speculative Decoding’s Potential for Accelerating Sparse MoE | — |
paper-2505-22179 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | quantization, speculative-decoding |
2505.22179 |
Speculative Decoding Meets Quantization: Compatibility Evaluation and Hierarchical Framework Design | — |
paper-2505-22913 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-pruning, sparse-attention |
2505.22913 |
Mustafar: Promoting Unstructured Sparsity for KV Cache Pruning in LLM Inference | — |
paper-2505-24544 |
screened |
paper |
not-classified |
not-classified |
— | authors | cross-attention, speculative-decoding |
2505.24544 |
Cross-Attention Speculative Decoding | — |
paper-2506-01979 |
screened |
paper |
not-classified |
not-classified |
— | authors | speculative-decoding |
2506.01979 |
Speculative Decoding via Hybrid Drafting and Rollback-Aware Branch Parallelism | — |
paper-2506-10274 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-codec, benchmark, survey |
2506.10274 |
Discrete Audio Tokens: More Than a Survey! | — |
paper-2507-12197 |
screened |
paper |
not-classified |
not-classified |
— | authors | audio-codec, parallel-codebooks, tts |
2507.12197 |
Quantize More, Lose Less: Autoregressive Generation from Residually Quantized Speech Representations | — |
paper-2507-18897 |
screened |
paper |
not-classified |
not-classified |
— | authors | audio-codec |
2507.18897 |
HH-Codec: High Compression High-fidelity Discrete Neural Codec for Spoken Language Modeling | — |
paper-2508-02558 |
screened |
paper |
not-classified |
not-classified |
— | authors | diffusion-lm, sparse-attention |
2508.02558 |
Sparse-dLLM: Accelerating Diffusion LLMs with Dynamic Cache Eviction | — |
paper-2508-15601 |
screened |
paper |
not-classified |
not-classified |
— | authors | mixed-precision, serving |
2508.15601 |
Efficient Mixed-Precision Large Language Model Inference with TurboMind | — |
paper-2508-18224 |
screened |
paper |
not-classified |
not-classified |
— | authors | kernel, sparse-attention |
2508.18224 |
Flash Sparse Attention: An Alternative Efficient Implementation of Native Sparse Attention Kernel | — |
paper-2509-02020 |
screened |
paper |
not-classified |
not-classified |
— | authors | dialogue, streaming, tts |
2509.02020 |
FireRedTTS-2: Towards Long Conversational Speech Generation for Podcast and Chatbot | — |
paper-2509-04377 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-pruning, paged-cache |
2509.04377 |
PagedEviction: Structured Block-wise KV Cache Pruning for Efficient LLM Inference | — |
paper-2509-05165 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-compression |
2509.05165 |
KVCompose: Efficient Structured KV Cache Compression with Composite Tokens | — |
paper-2509-11815 |
screened |
paper |
not-classified |
not-classified |
— | authors | speculative-decoding, vlm |
2509.11815 |
SpecVLM: Fast Speculative Decoding in Vision-Language Models | — |
paper-2510-00636 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-compression |
2510.00636 |
Expected Attention: KV Cache Compression by Estimating Attention from Future Queries Distribution | — |
paper-2510-16841 |
screened |
paper |
not-classified |
not-classified |
— | authors | audio-codec, semantic-acoustic |
2510.16841 |
SAC: Neural Speech Codec with Semantic-Acoustic Dual-Stream Quantization | — |
paper-2512-12087 |
screened |
paper |
not-classified |
not-classified |
— | authors | sparse-attention |
2512.12087 |
BLASST: Dynamic Blocked Attention Sparsity via Softmax Thresholding | — |
paper-2601-11580 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | benchmark, serving, speculative-decoding |
2601.11580 |
Speculative Decoding: Performance or Illusion? | — |
paper-2601-15621 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | dual-track, streaming, tts |
2601.15621 |
Qwen3-TTS Technical Report | — |
paper-2601-17702 |
screened |
paper |
not-classified |
not-classified |
— | authors | long-context, retrieval-attention |
2601.17702 |
S3-Attention: Attention-Aligned Endogenous Retrieval for Memory-Bounded Long-Context Inference | — |
paper-2601-19139 |
synthesized |
paper |
primary_paper |
paper_only |
performance, serving_semantics |
authors | continuous-batching, mlx, prefix-cache, serving, vlm |
2601.19139 |
Native LLM and MLLM Inference at Scale on Apple Silicon | Used for vllm-mlx source-reported throughput and repeated-media cache ranges; not generalized beyond its benchmark setup. |
paper-2601-19278 |
screened |
paper |
not-classified |
not-classified |
— | authors | diffusion-drafter, speculative-decoding |
2601.19278 |
DART: Diffusion-Inspired Speculative Decoding for Fast LLM Inference | — |
paper-2601-20094 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-codec, on-device, quantization |
2601.20094 |
T-Mimi: A Transformer-based Mimi Decoder for Real-Time On-Phone TTS | — |
paper-2601-23174 |
screened |
paper |
not-classified |
not-classified |
— | authors | speech-tokenizer, variable-rate |
2601.23174 |
Beyond Fixed Frames: Dynamic Character-Aligned Speech Tokenization | — |
paper-2602-04683 |
screened |
paper |
not-classified |
not-classified |
— | authors | audio-codec, audio-language |
2602.04683 |
UniAudio 2.0: A Unified Audio Language Model with Text-Aligned Factorized Audio Tokenization | — |
paper-2602-06036 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | dflash, speculative-decoding |
2602.06036 |
DFlash: Block Diffusion for Flash Speculative Decoding | — |
paper-2602-06072 |
screened |
paper |
not-classified |
not-classified |
— | authors | attention, batching, kernel |
2602.06072 |
PackInfer: Compute- and I/O-Efficient Attention for Batched LLM Inference | — |
paper-2602-10934 |
synthesized |
paper |
not-classified |
not-classified |
— | authors | audio-codec, transformer |
2602.10934 |
MOSS-Audio-Tokenizer: Scaling Audio Tokenizers for Future Audio Foundation Models | — |
paper-2602-23200 |
screened |
paper |
not-classified |
not-classified |
— | authors | hardware-aware, kv-quantization |
2602.23200 |
InnerQ: Hardware-Aware Tuning-Free Quantization of KV Cache for Large Language Models | — |
paper-2603-08899 |
screened |
paper |
not-classified |
not-classified |
— | authors | speculative-decoding |
2603.08899 |
ConFu: Contemplate the Future for Better Speculative Sampling | — |
paper-2603-11053 |
screened |
paper |
not-classified |
not-classified |
— | authors | scaling-laws, speculative-decoding |
2603.11053 |
Speculative Decoding Scaling Laws: Throughput Optimization Made Simple | — |
paper-2604-04722 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-quantization, on-device |
2604.04722 |
Don't Waste Bits! Adaptive KV-Cache Quantization for Lightweight On-Device LLMs | — |
paper-2604-07815 |
screened |
paper |
not-classified |
not-classified |
— | authors | offload, sparse-attention |
2604.07815 |
AsyncTLS: Efficient Generative LLM Inference with Asynchronous Two-level Sparse Attention | — |
paper-2604-12438 |
screened |
paper |
not-classified |
not-classified |
— | authors | low-latency, streaming, tts |
2604.12438 |
An Ultra-Low Latency, End-to-End Streaming Speech Synthesis Architecture via Block-Wise Generation and Depth-Wise Codec Decoding | — |
paper-2604-14493 |
screened |
paper |
not-classified |
not-classified |
— | authors | asr, on-device, quantization |
2604.14493 |
Pushing the Limits of On-Device Streaming ASR: A Compact, High-Accuracy English Model for Low-Latency Inference | — |
paper-2604-17852 |
screened |
paper |
not-classified |
not-classified |
— | authors | audio-codec, multi-token, semantic |
2604.17852 |
LLM-Codec: Neural Audio Codec Meets Language Model Objectives | — |
paper-2604-19157 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-quantization, serving |
2604.19157 |
SAW-INT4: System-Aware 4-Bit KV-Cache Quantization for Real-World LLM Serving | — |
paper-2605-17757 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-quantization, rotation |
2605.17757 |
OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization | — |
paper-2605-18226 |
screened |
paper |
not-classified |
not-classified |
— | authors | long-context, memory |
2605.18226 |
Context Memorization for Efficient Long Context Generation | — |
paper-2605-30852 |
indexed |
paper |
not-classified |
not-classified |
— | authors | pipeline, speculative-decoding |
2605.30852 |
Speculative Pipeline Decoding: Higher-Accuracy and Zero-Bubble Speculation via Pipeline Parallelism | — |
paper-2606-09079 |
screened |
paper |
not-classified |
not-classified |
— | authors | long-context, sparse-attention |
2606.09079 |
FlashMemory-DeepSeek-V4: Lightning Index Ultra-Long Context via Lookahead Sparse Attention | — |
paper-2606-20474 |
screened |
paper |
not-classified |
not-classified |
— | authors | agent-serving, kv-quantization |
2606.20474 |
UltraQuant: 4-bit KV Caching for Context-Heavy Agents | — |
paper-2606-21842 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-cache, rag, side-channel |
2606.21842 |
Agent-Assisted Side-Channel Attacks on Non-Prefix KV Cache in RAG | — |
paper-2606-24033 |
screened |
paper |
not-classified |
not-classified |
— | authors | kv-quantization, rope |
2606.24033 |
RoPE-Aware Bit Allocation for KV-Cache Quantization | — |
prompt-lookup-decoding-repo |
screened |
repository |
not-classified |
not-classified |
— | Apoorv Saxena | prompt-lookup, speculative-decoding |
2026-06-24 |
Prompt Lookup Decoding repository | Primary PLD source; not an MLX implementation by itself. |
pypi-mlx-audio |
screened |
release |
not-classified |
not-classified |
— | Blaizzy / contributors | asr, dependencies, mlx-audio, package, tts |
2026-06-27 |
mlx-audio PyPI package metadata | Package metadata evidence for audio dependency and model-matrix review gates; validate per model. |
pypi-mlx-lm |
screened |
release |
not-classified |
not-classified |
— | ml-explore | cache, mlx-lm, package, serving |
2026-06-27 |
mlx-lm PyPI package metadata | Package metadata evidence for release/version and prompt-cache caveats; PyPI text alone is not implementation proof. |
pypi-mlx-vlm |
screened |
release |
not-classified |
not-classified |
— | Blaizzy / contributors | cache, mlx-vlm, package, speculative-decoding, vlm |
2026-06-27 |
mlx-vlm PyPI package metadata | Package metadata evidence for feature-cache, speculative decoding, and TurboQuant leads; validate against pinned source before promotion. |
pytorch-repo |
indexed |
repository |
not-classified |
not-classified |
— | PyTorch | operators, source-oracle |
— |
PyTorch repository | — |
repo-audiocraft |
screened |
repository |
not-classified |
not-classified |
— | Meta | audiogen, encodec, musicgen |
— |
AudioCraft repository | — |
repo-bigvgan |
screened |
repository |
not-classified |
not-classified |
— | NVIDIA | vocoder |
— |
BigVGAN repository | — |
repo-dac |
screened |
repository |
not-classified |
not-classified |
— | Descript | audio-codec |
— |
Descript Audio Codec repository | — |
repo-encodec |
screened |
repository |
not-classified |
not-classified |
— | Meta | audio-codec |
— |
EnCodec repository | — |
repo-f5tts |
screened |
repository |
not-classified |
not-classified |
— | SWivid | flow-matching, tts |
— |
F5-TTS repository | — |
repo-fish-speech |
screened |
repository |
not-classified |
not-classified |
— | Fish Audio | audio-lm, codec, tts |
— |
Fish Speech repository | — |
repo-funcodec |
screened |
repository |
not-classified |
not-classified |
— | Alibaba | audio-codec |
— |
FunCodec repository | — |
repo-moonshine |
screened |
repository |
not-classified |
not-classified |
— | Useful Sensors | asr, on-device |
— |
Moonshine repository | — |
repo-moshi |
screened |
repository |
not-classified |
not-classified |
— | Kyutai | codec, duplex, streaming-speech |
— |
Moshi repository | — |
repo-nemo |
screened |
repository |
not-classified |
not-classified |
— | NVIDIA | asr, canary, parakeet, tts |
— |
NVIDIA NeMo speech repository | — |
repo-qwen-audio |
screened |
repository |
not-classified |
not-classified |
— | Qwen | audio-language |
— |
Qwen Audio repository | — |
repo-qwen-omni |
screened |
repository |
not-classified |
not-classified |
— | Qwen | omni, speech |
— |
Qwen2.5-Omni repository | — |
repo-stable-audio-tools |
indexed |
repository |
not-classified |
not-classified |
— | Stability AI | audio-generation, diffusion |
— |
Stable Audio Tools repository | — |
repo-vocos |
screened |
repository |
not-classified |
not-classified |
— | gemelo-ai | fourier, vocoder |
— |
Vocos repository | — |
repo-whisper |
screened |
repository |
not-classified |
not-classified |
— | OpenAI | asr, whisper |
— |
Whisper repository | — |
safetensors-repo |
screened |
repository |
not-classified |
not-classified |
— | Hugging Face | security, serialization, weights |
— |
Safetensors format repository | — |
skills-ref-repo |
screened |
repository |
not-classified |
not-classified |
— | Agent Skills | agent-skills, validation |
5d4c1fda3f786fff826c7f56b6cb3341e7f3a911 |
Agent Skills reference validator | — |
skryl-mlx-onnx-repo |
screened |
repository |
not-classified |
not-classified |
— | skryl | export, graph-ir, mlx, onnx |
c9aec867c76b |
mlx-onnx export library | Screened as top-1000 contributor context; export tooling is useful adjacent work but not a model-to-MLX porting technique. |
skyzh-tiny-llm-repo |
screened |
repository |
not-classified |
not-classified |
— | skyzh | continuous-batching, gather-qmm, mlx, moe, paged-kv |
efb0c89fd236 |
tiny-llm MLX serving course repository | Screened as top-1000 contributor implementation context; reinforces existing paged-cache and MoE validation gates. |
spheron-nvfp4-mxfp4-guide |
screened |
technical-blog |
not-classified |
not-classified |
— | Spheron | low-bit-formats, mxfp4, nvfp4, quantization |
2026-06-24 |
NVFP4 vs MXFP4: 4-Bit Quantization Format Decision Guide | Practitioner format comparison; use official MLX docs for supported modes. |
tristanbilot-mlx-benchmark-repo |
screened |
repository |
not-classified |
not-classified |
— | TristanBilot | benchmarking, cuda, mlx, mps, mx.compile |
fc7b2fa714bc |
mlx-benchmark operation benchmark repository | Screened as top-100 contributor benchmarking context; useful measurement hygiene, not a port-level optimization claim. |
vllm-blog-anatomy |
screened |
technical-blog |
not-classified |
not-classified |
— | vLLM project | continuous-batching, paged-attention, prefix-cache, serving, speculative-decoding |
2026-06-24 |
Inside vLLM: Anatomy of a High-Throughput LLM Inference System | Serving-system practitioner context; not MLX-specific implementation evidence. |
vllm-doc-prefix-caching |
screened |
official-doc |
not-classified |
not-classified |
— | vLLM project | kv-cache, prefix-cache, serving |
2026-06-24 |
vLLM automatic prefix caching documentation | General serving reference for prefix-cache behavior; not MLX framework support. |
vllm-mlx-continuous-batching-guide |
synthesized |
source-code |
primary_source_code |
third_party_pinned |
mlx_implementation, performance, serving_semantics |
waybarrios | continuous-batching, serving, vllm-mlx |
a48c86c1a41900f7d26658471b5f67e5fdd35445 |
vLLM-MLX continuous batching guide | Third-party pinned serving guide; text/VLM evidence must not be promoted to generic audio support. |
vllm-mlx-memory-cache-source |
synthesized |
source-code |
not-classified |
not-classified |
— | waybarrios | memory, prefix-cache, serving, vllm-mlx |
a48c86c1a41900f7d26658471b5f67e5fdd35445 |
vLLM-MLX memory-aware prefix cache implementation | Pinned evidence for memory accounting and cache constraints in third-party serving. |
vllm-mlx-mllm-batch-source |
synthesized |
source-code |
primary_source_code |
third_party_pinned |
mlx_implementation, serving_semantics |
waybarrios | batching, multimodal, prefix-cache, vllm-mlx |
a48c86c1a41900f7d26658471b5f67e5fdd35445 |
vLLM-MLX multimodal batch generator implementation | Implementation evidence for multimodal cache integration and media-placeholder safety behavior. |
vllm-mlx-multimodal-guide |
synthesized |
source-code |
primary_source_code |
third_party_pinned |
mlx_implementation, serving_semantics |
waybarrios | multimodal, serving, video, vllm-mlx |
a48c86c1a41900f7d26658471b5f67e5fdd35445 |
vLLM-MLX multimodal serving guide | Source for video FPS/max-frame knobs and multimodal serving caveats. |
vllm-mlx-repo |
screened |
repository |
not-classified |
not-classified |
— | waybarrios | continuous-batching, paged-kv, prefix-cache, serving |
a48c86c1a41900f7d26658471b5f67e5fdd35445 |
vLLM-MLX serving repository | — |
vllm-mlx-vision-cache-source |
synthesized |
source-code |
primary_source_code |
third_party_pinned |
mlx_implementation, serving_semantics |
waybarrios | multimodal, serving, vision-cache, vllm-mlx |
a48c86c1a41900f7d26658471b5f67e5fdd35445 |
vLLM-MLX vision embedding cache implementation | Two-level media preprocessing/vision-encoding cache implementation. |