From 6e7c0a4df815d43e090e80055bc6c86da4aa99ec Mon Sep 17 00:00:00 2001 From: "mudler's LocalAI [bot]" <139863280+localai-bot@users.noreply.github.com> Date: Mon, 3 Aug 2026 23:03:25 +0200 Subject: [PATCH] blog, website: edit out the AI writing tells readers called out on HN (#11324) * blog: rewrite the engines post without the AI tells The HN thread on this post (item 49125065) spent most of its comments on the writing rather than the engines. Readers quoted specific lines back as tells. This is the same post with the same numbers, edited against the updated no-ai-slop skill. Every figure, table and link is unchanged, except that "27% of the memory" is now the underlying 363 MB against 1328 MB from the table. Two substantive framing fixes, both from the reply draft in hn-reply-engines-post.md: - vllm.cpp is no longer implied to be a speed win. The table is a tie, the result is the install size, and the post now says so before a reader has to work it out and post about it. - Added one line on the language mix. Readers took the C++/Python/Go tree as incoherence rather than as a Go core with per-ecosystem backends. Cut throughout: the ledger metaphor ("what those ports buy", "not paid for in throughput"), unearned framing ("the honest reading is", "has nothing to do with"), the shape summary ("that is the general shape of these wins"), confident deference ("people who are better at those models than we are"), self-grading numbers ("a good result for a 66 MiB binary"), verbless comparisons, three of the four exactness idioms, and the aphoristic headings and verdicts. The double-tricolon summary is one plain clause now. Co-Authored-By: Claude Opus 5 (1M context) * blog, website: same anti-slop sweep over the rest of the site One-by-one pass over the other four posts and the site templates, with the same rules used on the engines post. All figures, tables, links and PR numbers are unchanged everywhere; the edits are to prose only. apex-moe-quantization: ledger metaphors were the main issue, eight uses of buy/cost/pay/spend for things that are not money. Also "the honest reading is", "that is the comparison that matters", and two section-ending aphorisms ("Size is a speed knob as much as a memory knob", "Q6_K is the ceiling worth paying for"). localai-since-march-2023: light touch, this one already reads like a person. Removed "the curve is not the point", a "not the feature list, but the four decisions" contrast, and two "X is what made / is the piece that" forms. parakeet-cpp-asr-on-cpu: six exactness idioms across one post, "byte for byte" twice, "character for character" twice, "byte-identical" twice and "bit-identical" once, including in the title. Down to one, kept where the precision is load-bearing. Also the "what end-of-utterance detection buys you" heading and the "we say so rather than averaging it away" flex. what-landed-in-localai-4-8: no changes. It is dense, flat and ends every section on a PR number or a plain fact, which is the shape the other posts should look like. Site templates: "Most backends wrap somebody else's engine. These do not." was the same contrast the engines post opened with. Also "Not a degraded mode that technically runs", "A port only ships once it matches the original", "Speed is the part we then go and win ... not a marketing run", and the last "byte for byte" on the landing page. Hugo builds clean. Co-Authored-By: Claude Opus 5 (1M context) * website: it is eighteen engines, not nineteen Three places said nineteen: the /engines/ page description, the JUL 2026 timeline entry on the landing page, and the header comment in data/engines.yaml. Eighteen is right, confirmed two ways. The "Backends built by us" table in the README has exactly 18 rows, and data/engines.yaml has 19 entries of which one is apex-quant, which is a quantization recipe rather than an engine. The two lists otherwise match name for name. The yaml comment is the likely origin: it read "the nineteen native engines the LocalAI team wrote, and the one quantization recipe that feeds them", which counts apex-quant twice. Co-Authored-By: Claude Opus 5 (1M context) --------- Co-authored-by: Ettore Di Giacinto Co-authored-by: Claude Opus 5 (1M context) --- website/content/blog/_index.md | 2 +- website/content/blog/apex-moe-quantization.md | 26 ++++----- .../content/blog/localai-since-march-2023.md | 12 ++-- .../content/blog/parakeet-cpp-asr-on-cpu.md | 24 ++++---- .../blog/why-we-write-our-own-engines.md | 58 ++++++++++--------- website/content/engines/_index.md | 2 +- website/data/engines.yaml | 2 +- website/layouts/engines/list.html | 6 +- website/layouts/index.html | 10 ++-- 9 files changed, 73 insertions(+), 69 deletions(-) diff --git a/website/content/blog/_index.md b/website/content/blog/_index.md index 8c95e2e38..f921a308c 100644 --- a/website/content/blog/_index.md +++ b/website/content/blog/_index.md @@ -1,5 +1,5 @@ --- title: "Blog" -description: "Release write-ups, benchmark reports and engineering notes from the LocalAI team. Every number here comes out of a benchmark suite, a release or a commit, and the source is named so you can check it." +description: "Release write-ups, benchmark reports and engineering notes from the LocalAI team. Numbers link to the release, commit or benchmark run they came from." extracss: ["blog.css"] --- diff --git a/website/content/blog/apex-moe-quantization.md b/website/content/blog/apex-moe-quantization.md index b7de05d5f..4a904dd14 100644 --- a/website/content/blog/apex-moe-quantization.md +++ b/website/content/blog/apex-moe-quantization.md @@ -4,13 +4,13 @@ date: 2026-04-10 author: "Ettore Di Giacinto" category: "Research" tags: ["quantization", "APEX", "mixture-of-experts", "llama.cpp", "benchmarks"] -summary: "Qwen3.5-35B-A3B goes from 64.6 GB to 12.2 GB and speeds up from 30.4 to 74.4 tokens per second. Perplexity moves from 6.537 to 7.088. Here is the precision assignment that does it, and where it costs you." +summary: "Qwen3.5-35B-A3B goes from 64.6 GB to 12.2 GB and speeds up from 30.4 to 74.4 tokens per second. Perplexity moves from 6.537 to 7.088. Here is the precision assignment that does it, and where the quality drops." extracss: ["blog.css"] --- A 35B mixture-of-experts model at full precision is a 64.6 GB file, which puts it out of reach of every consumer GPU. APEX gets Qwen3.5-35B-A3B down to 12.2 GB, where it fits a 16 GB card with room for context, and it generates at 74.4 tokens per second instead of 30.4. The output is an ordinary GGUF that stock llama.cpp opens with no patches and no custom build. -The compression is not free at that tier, and the numbers below say exactly what it costs. At the 21.3 GB tier it is closer to free than we expected: APEX Quality has a lower perplexity than the F16 model it was quantized from. +At that tier the quality does drop, and the numbers below say by how much. At the 21.3 GB tier it barely drops at all: APEX Quality has a lower perplexity than the F16 model it was quantized from. ## The measurements @@ -37,17 +37,17 @@ All of this is Qwen3.5-35B-A3B on an NVIDIA DGX Spark (GB10, 122 GB unified VRAM Three things in that table are worth stopping on. -APEX Quality is 21.3 GB, a third of F16, and its perplexity of 6.527 is lower than F16's 6.537 and lower than Q8_0's 6.533. Quantization noise acting as mild regularization on a wikitext evaluation is a known effect and we are not claiming the quantized model is smarter. The honest reading is that at this tier the loss is below the measurement floor. +APEX Quality is 21.3 GB, a third of F16, and its perplexity of 6.527 is lower than F16's 6.537 and lower than Q8_0's 6.533. Quantization noise acting as mild regularization on a wikitext evaluation is a known effect and we are not claiming the quantized model is smarter. At this tier the loss is below the measurement floor. -Against Unsloth's UD-Q8_K_XL, APEX I-Quality is half the size (21.3 GB against 45.3 GB), one point ahead on HellaSwag (83.5% against 82.5%), within 0.016 on perplexity, and 73% faster (63.1 t/s against 36.4). That is the comparison that matters for anyone choosing a published quant today. +Against Unsloth's UD-Q8_K_XL, APEX I-Quality is half the size (21.3 GB against 45.3 GB), one point ahead on HellaSwag (83.5% against 82.5%), within 0.016 on perplexity, and 73% faster (63.1 t/s against 36.4). At the bottom end, APEX Mini beats bartowski IQ2_M on every metric while being 0.9 GB larger: perplexity 7.088 against 7.303, HellaSwag 81.0% against 80.3%, MMLU 41.3% against 39.6%. -## Why it gets faster, not just smaller +## Why it also gets faster Token generation on a single stream is bound by memory bandwidth, not by arithmetic. Every generated token requires reading the active weights out of memory, so halving the bytes roughly halves the time spent waiting for them. Going from 64.6 GB to 12.2 GB takes throughput from 30.4 to 74.4 tokens per second, a 2.45x gain on the same hardware with the same kernels. Every APEX tier clears 60 t/s. -That is also why a large well-behaved quant such as UD-Q8_K_XL is slower than a smaller one with equal quality. Size is a speed knob as much as a memory knob. +That is also why a large well-behaved quant such as UD-Q8_K_XL is slower than a smaller one with equal quality. ## Per-tensor and per-layer precision @@ -55,9 +55,9 @@ Uniform quantization gives every tensor the same bit width, which spends the sam APEX classifies every tensor into one of three roles and treats them differently. -**Routed expert weights** (the gate, up and down projections inside the experts) are the bulk of the parameters, and only 8 of 256 experts are active per token. That 97% structural sparsity is what makes aggressive quantization safe here. The routing decision itself reads full-precision gate weights, so quantization noise inside an expert that was not selected never reaches the output at all. When an expert is selected, its contribution is one of eight summed paths, which further dilutes per-tensor error. +**Routed expert weights** (the gate, up and down projections inside the experts) are the bulk of the parameters, and only 8 of 256 experts are active per token. That 97% structural sparsity is why aggressive quantization is safe here. The routing decision itself reads full-precision gate weights, so quantization noise inside an expert that was not selected never reaches the output at all. When an expert is selected, its contribution is one of eight summed paths, which further dilutes per-tensor error. -**Shared expert weights** run for every single token and their weight distribution is heavy-tailed, with a kurtosis of 13.10 against 3.41 for routed experts. Those outliers carry real signal and low-bit formats clip them. Q8_0 is the minimum viable precision here, and dropping it is the fastest way to wreck a build. +**Shared expert weights** run for every single token and their weight distribution is heavy-tailed, with a kurtosis of 13.10 against 3.41 for routed experts. Those outliers carry real signal and low-bit formats clip them. Q8_0 is the minimum viable precision here, and dropping it degrades the build quickly. **Attention and SSM weights** are dense, contribute few parameters relative to the experts, and matter for generation quality. They sit at Q6_K throughout. @@ -69,23 +69,23 @@ None of this needs a patched llama.cpp. The assignments are expressed with the s Twenty-five or so systematic runs produced a few results that saved a lot of time later. -Going from Q6_K to Q8_0 on routed experts costs 7.5 GB and buys zero perplexity improvement. Going below Q5_K on them causes measurable degradation. Q6_K is the ceiling worth paying for. +Going from Q6_K to Q8_0 on routed experts costs 7.5 GB and gives zero perplexity improvement. Going below Q5_K on them causes measurable degradation. Q6_K is the ceiling. Layer position matters more than uniform bit width. A two-tier gradient of Q6_K edges and Q5_K middle matches Q8_0 quality; a uniform Q5_K assignment at a similar size does not. IQ formats underperform K-quants on MoE experts. IQ3_S gives worse perplexity than Q3_K on routed expert tensors at a similar bit rate, because the near-Gaussian expert weight distribution (kurtosis 3.41) suits the K-quant block structure better. -Five C-level modifications to the quantization algorithms themselves, including error feedback, enhanced scale search, super-block refinement and Gaussian-density weighting, all showed zero improvement. Stock llama.cpp quantization is already good. The gains here come entirely from deciding where to spend bits. +Five C-level modifications to the quantization algorithms themselves, including error feedback, enhanced scale search, super-block refinement and Gaussian-density weighting, all showed zero improvement. Stock llama.cpp quantization is already good. The gains here come entirely from deciding where to put the bits. ## The I-variants and their calibration set Standard imatrix calibration uses Wikipedia text, which is also what wikitext perplexity measures, so the calibration and the benchmark agree with each other by construction. The I-variants calibrate on a diverse set spanning chat, code, reasoning and tool-calling, with no Wikipedia in it. -That trade shows up clearly. I-Compact drops perplexity from 6.783 to 6.669, cuts KL max from 7.56 to 5.50, and lifts MMLU from 40.9% to 41.7%. At the Quality tier, I-Quality gives up 0.025 perplexity against Quality and takes the highest HellaSwag score of anything tested (83.5%), the best TruthfulQA (38.4%), and a lower KL divergence. If your workload is chat, code or agents rather than encyclopedic prose, take the I variant. +It shows up in the numbers. I-Compact drops perplexity from 6.783 to 6.669, cuts KL max from 7.56 to 5.50, and lifts MMLU from 40.9% to 41.7%. At the Quality tier, I-Quality gives up 0.025 perplexity against Quality and takes the highest HellaSwag score of anything tested (83.5%), the best TruthfulQA (38.4%), and a lower KL divergence. If your workload is chat, code or agents rather than encyclopedic prose, take the I variant. -## Where it costs you +## Where the quality drops -The Compact and Mini tiers are real compression, and they are not free. +The Compact and Mini tiers lose real quality. Compact at 16.1 GB moves perplexity from 6.537 to 6.783, a 3.8% increase, and its KL mean rises tenfold against Q8_0, from 0.0046 to 0.0469. Mini at 12.2 GB goes to 7.088, an 8.4% increase, with a KL mean of 0.0870 and HellaSwag down 1.5 points to 81.0%. Those are the numbers to weigh against the fact that the model now runs at all on a 16 GB card. diff --git a/website/content/blog/localai-since-march-2023.md b/website/content/blog/localai-since-march-2023.md index 9f2782918..9ad7109da 100644 --- a/website/content/blog/localai-since-march-2023.md +++ b/website/content/blog/localai-since-march-2023.md @@ -4,7 +4,7 @@ date: 2026-07-29 author: "Ettore Di Giacinto" category: "History" tags: ["history", "architecture", "releases", "community"] -summary: "Three years, 133 releases and 224 contributors later. The four changes that mattered most were making the core small, adding agents, making it a cluster, and giving it eyes and ears." +summary: "Three years, 133 releases and 224 contributors later. Here are the four decisions that shaped it: making the core small, adding agents, making it a cluster, and giving it eyes and ears." extracss: ["blog.css"] --- @@ -16,9 +16,9 @@ None of those numbers are rounded up. You can read every one of them off the rep {{< starchart >}} -The curve is not the point, but it is a useful map. The four marks on it are the four decisions below, and you can see each of them in the slope afterwards. +The four marks on it are the four decisions below, and you can see each of them in the slope afterwards. -What follows is how it got here. Not the feature list, which you can read in the releases, but the four decisions that changed the shape of the thing. +What follows is the four decisions that changed the shape of the thing. The full feature list is in the releases. ## 2023 to 2024: an API in front of llama.cpp @@ -34,7 +34,7 @@ Every backend moved out of the main binary in [v3.2.0](https://github.com/mudler You install one thing and it stays small. Ask for a GGUF model and llama-cpp arrives. Ask for transcription and whisper or parakeet arrives. Nothing else is fetched, and a machine that only ever serves one model never downloads the other sixty-nine backends. -That one change is what made everything after it possible. Adding a backend stopped meaning adding weight to everybody's install, so "should we support this engine" stopped being an argument about download size and went back to being an argument about whether the engine is any good. It is also the reason we can afford to maintain eighteen engines of our own, which comes later. +Everything after it depended on that one change. Adding a backend stopped meaning adding weight to everybody's install, so "should we support this engine" stopped being an argument about download size and went back to being an argument about whether the engine is any good. It is also the reason we can afford to maintain eighteen engines of our own, which comes later. ## March 2026: agents, and a new interface @@ -42,7 +42,7 @@ That one change is what made everything after it possible. Adding a backend stop The web interface was rewritten in React at the same time, with a Canvas mode, MCP Apps and client-side tools with tool streaming ([#8947](https://github.com/mudler/LocalAI/pull/8947)), and WebRTC realtime audio ([#8790](https://github.com/mudler/LocalAI/pull/8790)). MLX gained a distributed mode ([#8801](https://github.com/mudler/LocalAI/pull/8801)). -The realtime audio path is the piece that changed what people built. Speech in, tool calls in the middle, speech out, over WebRTC, fast enough that it feels like a conversation rather than a walkie-talkie. It had landed as the Realtime API in February 2026 ([#6245](https://github.com/mudler/LocalAI/pull/6245)), and the interface rewrite finally gave it a face. +The realtime audio path changed what people built with it. Speech in, tool calls in the middle, speech out, over WebRTC, fast enough that it feels like a conversation rather than a walkie-talkie. It had landed as the Realtime API in February 2026 ([#6245](https://github.com/mudler/LocalAI/pull/6245)), and the interface rewrite finally gave it a face. ## April 2026: it becomes a cluster @@ -74,6 +74,6 @@ The most recent one is [vllm.cpp](https://github.com/mudler/vllm.cpp), a C++20 p ## Where it stands -Still MIT, still a community project. 224 people have put code in, and the README is kept translated into eight languages because the people using this are not all in one place. The [contributors graph](https://github.com/mudler/LocalAI/graphs/contributors) is the honest picture of who actually built this, and it is not me. +Still MIT, still a community project. 224 people have put code in, and the README is kept translated into eight languages because the people using this are not all in one place. The [contributors graph](https://github.com/mudler/LocalAI/graphs/contributors) shows who actually built this, and it is not me. If you want to add something, backends and gallery entries are the two places a first contribution lands cleanly. There is a step-by-step checklist for a new backend in `.agents/adding-backends.md`, and a gallery entry is just a YAML block. Come say hello in [Discord](https://discord.gg/uJAeKSAGDy) if you get stuck. diff --git a/website/content/blog/parakeet-cpp-asr-on-cpu.md b/website/content/blog/parakeet-cpp-asr-on-cpu.md index 2942c888c..7352a4edd 100644 --- a/website/content/blog/parakeet-cpp-asr-on-cpu.md +++ b/website/content/blog/parakeet-cpp-asr-on-cpu.md @@ -1,22 +1,22 @@ --- -title: "parakeet.cpp: NeMo transcripts, byte for byte, without the Python" +title: "parakeet.cpp: the same NeMo transcript, without the Python" date: 2026-06-05 author: "Ettore Di Giacinto" category: "Benchmarks" tags: ["parakeet.cpp", "ASR", "ggml", "streaming", "benchmarks"] -summary: "Same transcript as NVIDIA NeMo, character for character, at a median 1.40x on CPU and about 27x the speed of whisper.cpp. One binary, one GGUF file, no Python at inference." +summary: "The same transcript as NVIDIA NeMo at a median 1.40x on CPU, and about 27x the speed of whisper.cpp, from one binary and one GGUF file." extracss: ["blog.css"] --- -You can drop a single binary and a GGUF file onto a machine with no GPU and get NVIDIA NeMo Parakeet transcription out of it, at a median 1.40x NeMo's own PyTorch CPU speed, with a transcript that matches NeMo character for character. That is [parakeet.cpp](https://github.com/mudler/parakeet.cpp), a C++17 port of the Parakeet speech-recognition family built on ggml. +You can drop a single binary and a GGUF file onto a machine with no GPU and get NVIDIA NeMo Parakeet transcription out of it, at a median 1.40x NeMo's own PyTorch CPU speed, with the same transcript NeMo produces. That is [parakeet.cpp](https://github.com/mudler/parakeet.cpp), a C++17 port of the Parakeet speech-recognition family built on ggml. -Accuracy came first and speed came second, in that order, because a faster transcriber that disagrees with the reference is a different model, not a port. +We checked the accuracy before touching the speed, because a transcriber that disagrees with the reference is not a port of it. ## WER 0 against NeMo Every published checkpoint is validated at WER 0 against NeMo. Across the LibriSpeech test-clean set the mean f32 agreement WER, meaning the word error rate between our transcript and NeMo's on the same audio, is 0.0155%. On seven of the ten models it is exactly 0.0000%, which is a byte-identical transcript. -That number is what makes the speed comparison meaningful. Both engines did the same work and produced the same output, so the only difference left is how long they took. +Both engines did the same work and produced the same output, so the only difference left is how long they took. ## CPU, against NeMo's own runtime @@ -54,25 +54,25 @@ Against whisper.cpp turbo on the same clip and at the same accuracy (1.6% WER on The decisive win was on the decode side. A transducer decodes autoregressively, and profiling showed the prediction-network LSTM taking about 97% of RNN-T decode time while producing the same output over and over: on a non-emitting frame the prediction network's input has not changed, so its forward pass is redundant. Caching that forward across non-emitting frames removed most of the decode cost. -The encoder side is a set of smaller wins with no single hero: a persistent ggml backend with `gallocr`, zero-copy weights straight out of the GGUF mapping, one fused graph rather than per-layer graph building, and tinyBLAS through `GGML_LLAMAFILE`. +The encoder side is a set of smaller wins: a persistent ggml backend with `gallocr`, zero-copy weights straight out of the GGUF mapping, one fused graph rather than per-layer graph building, and tinyBLAS through `GGML_LLAMAFILE`. ## On the GPU On an NVIDIA GB10 (Grace-Blackwell), parakeet.cpp wins on all ten models, with a median of 1.25x and up to 4.3x on the large TDT and hybrid models. The reference here is NeMo-GPU inside the `nvcr.io/nvidia/nemo` container, because NeMo cannot run on that host's torch and CUDA stack directly. -The 4.3x cases have a specific cause. NeMo's TDT greedy decode is not CUDA-graph accelerated and falls back to a per-step Python loop, while ours is a lean C++ loop. Where NeMo's decode is CUDA-graph accelerated, as it is for RNN-T, the gap narrows to about 1.16x at f32 and 1.30x at q8_0. On the pure-encoder CTC models the margin is around 1.2x, because ggml's generic CUDA conv and attention kernels still trail NVIDIA's tuned cuDNN. That is the main piece of GPU headroom left in the project and we say so in the README rather than averaging it away. +The 4.3x cases have a specific cause. NeMo's TDT greedy decode is not CUDA-graph accelerated and falls back to a per-step Python loop, while ours is a lean C++ loop. Where NeMo's decode is CUDA-graph accelerated, as it is for RNN-T, the gap narrows to about 1.16x at f32 and 1.30x at q8_0. On the pure-encoder CTC models the margin is around 1.2x, because ggml's generic CUDA conv and attention kernels still trail NVIDIA's tuned cuDNN. That is the main piece of GPU headroom left in the project, and the README lists it per model. -Batching several clips through the decoder together reaches about 10x to 12x at batch size 16 on the GB10, and about 3x to 5x on CPU. It applies to transducer models only, since CTC has no autoregressive decode to batch, and the batched path is bit-identical to running the clips one at a time. +Batching several clips through the decoder together reaches about 10x to 12x at batch size 16 on the GB10, and about 3x to 5x on CPU. It applies to transducer models only, since CTC has no autoregressive decode to batch, and the batched path produces the same output as running the clips one at a time. On Apple M4 through ggml's Metal backend, the larger models run about 3x to 5x faster than the same models on that machine's CPU. -## Cache-aware streaming, and what end-of-utterance detection buys you +## Cache-aware streaming and end-of-utterance detection Offline transcription hands you a file and waits. A voice assistant cannot do that, so `parakeet_realtime_eou_120m-v1` runs a cache-aware streaming path instead: you feed it 16 kHz mono PCM as it arrives and it returns newly finalized text as it becomes stable. -Cache-aware means the cost per chunk stays flat. Each chunk's forward pass carries per-layer convolution and attention caches plus the transducer decoder state forward, so nothing before the current chunk is recomputed. Without that, every chunk would re-run the encoder over the whole session so far, and the per-chunk cost would grow with the length of the conversation until the loop fell behind. The implementation covers layer norm with causal convolution, causal subsampling, and chunked-limited attention, and its transcript matches NeMo's own cache-aware streaming byte for byte. +Cache-aware means the cost per chunk stays flat. Each chunk's forward pass carries per-layer convolution and attention caches plus the transducer decoder state forward, so nothing before the current chunk is recomputed. Without that, every chunk would re-run the encoder over the whole session so far, and the per-chunk cost would grow with the length of the conversation until the loop fell behind. The implementation covers layer norm with causal convolution, causal subsampling, and chunked-limited attention, and its transcript matches NeMo's own cache-aware streaming exactly. -End-of-utterance detection is the part that changes how an assistant feels. The model emits `` when the speaker has finished a turn and `` for a backchannel, as events alongside the text. A voice loop can start generating a reply the moment `` arrives rather than waiting out a fixed silence timer, which is where most of the perceived lag in a spoken assistant comes from. The alternative, a VAD with a 700 ms hangover, either cuts people off mid-sentence or makes the assistant feel slow, and it cannot tell "mm-hm" from the end of a thought. `finalize` flushes the tail at end of stream without fabricating an `` that NeMo would not have emitted. +End-of-utterance detection changes how an assistant feels. The model emits `` when the speaker has finished a turn and `` for a backchannel, as events alongside the text. A voice loop can start generating a reply the moment `` arrives rather than waiting out a fixed silence timer, which is where most of the perceived lag in a spoken assistant comes from. The alternative, a VAD with a 700 ms hangover, either cuts people off mid-sentence or makes the assistant feel slow, and it cannot tell "mm-hm" from the end of a thought. `finalize` flushes the tail at end of stream without fabricating an `` that NeMo would not have emitted. The streaming path measures at RTFx 3.80 on a 7.43 second clip. That sits well below the offline number by design, because streaming runs many small chunked passes rather than one large one, and it is still several times faster than real time on a CPU. @@ -95,7 +95,7 @@ parakeet.cpp ports NeMo's `rel_pos_local_attn`, a banded attention where each qu -At NeMo's full W=128 window that is about 4x faster and about 5.7x less peak memory than the global path. The band is built with a chunk-matmul construction, overlapping key and value chunks feeding one batched GEMM plus a diagonal skew view, so the graph node count does not depend on the window. The wide window costs the same as the narrow one. Short clips stay on the global path and remain byte-identical to before. +At NeMo's full W=128 window that is about 4x faster and about 5.7x less peak memory than the global path. The band is built with a chunk-matmul construction, overlapping key and value chunks feeding one batched GEMM plus a diagonal skew view, so the graph node count does not depend on the window. The wide window costs the same as the narrow one. Short clips stay on the global path and produce the same output as before. ## Using it diff --git a/website/content/blog/why-we-write-our-own-engines.md b/website/content/blog/why-we-write-our-own-engines.md index 411242b08..2d09cc0c7 100644 --- a/website/content/blog/why-we-write-our-own-engines.md +++ b/website/content/blog/why-we-write-our-own-engines.md @@ -4,21 +4,21 @@ date: 2026-07-24 author: "Ettore Di Giacinto" category: "Engineering" tags: ["engineering", "ggml", "vllm.cpp", "depth-anything.cpp", "parity"] -summary: "A 66 MiB binary instead of a 9.1 GiB virtualenv, depth estimation that beats PyTorch on CPU in half the memory, and biometrics that match insightface bit for bit. The method, the measurements, and what it costs us." +summary: "Eighteen of our backends are C or C++ ports we wrote from scratch instead of wrapping an upstream engine. Here is why we did it and what we measured." extracss: ["blog.css"] --- -Most LocalAI backends wrap somebody else's engine, and that is the right default. llama.cpp, vLLM, whisper.cpp, stable-diffusion, MLX and the rest are maintained by people who are better at those models than we are, and wrapping them costs a Dockerfile and a gRPC shim. +Most LocalAI backends wrap somebody else's engine. llama.cpp, vLLM, whisper.cpp, stable-diffusion and MLX are maintained by people who work on those models full time, and wrapping one of them costs us a Dockerfile and a gRPC shim. We do that wherever we can. -Eighteen of our backends do not wrap anything. They are C or C++ ports we wrote from scratch, and each one exists because wrapping the upstream engine would have meant shipping something we could not ship: a multi-gigabyte Python install, a non-portable CUDA-only stack, or a model that had no C++ implementation at all. This post is about what those ports buy, measured, and what they cost. +Eighteen of our backends do not wrap anything. They are C or C++ ports we wrote from scratch, and each one exists because wrapping the upstream engine would have meant shipping something we could not ship: a multi-gigabyte Python install, a CUDA-only stack that will not run on half the machines our users have, or, in a few cases, a model with no C++ implementation to wrap in the first place. Below are the numbers for four of them, and what keeping them alive takes. -## What you get: one file, and memory you can predict +## vllm.cpp: 66 MiB instead of 9.1 GiB -Deploying a Python inference stack means resolving a dependency tree at install time, on the target machine, against whatever CUDA and glibc it has. Deploying a ggml port means copying a shared library and a GGUF file. +Deploying a Python inference stack means resolving a dependency tree at install time, on the target machine, against whatever CUDA and glibc that machine has. Deploying a ggml port means copying a shared library and a GGUF file. -The clearest measurement of that difference is [vllm.cpp](https://github.com/mudler/vllm.cpp), our C++20 port of vLLM's V1 serving architecture. Installing vLLM produces a 9.1 GiB virtualenv. Installing vllm.cpp produces a 66 MiB binary. The engine implements the same things the Python original does, including paged KV cache, continuous batching, prefix caching, the scheduler and the sampler, with no Python, no PyTorch and no ggml at inference. +[vllm.cpp](https://github.com/mudler/vllm.cpp) is our C++20 port of vLLM's V1 serving architecture. Installing vLLM produces a 9.1 GiB virtualenv. Installing vllm.cpp produces a 66 MiB binary. It implements the same things the Python original does, including paged KV cache, continuous batching, prefix caching, the scheduler and the sampler, with no Python, no PyTorch and no ggml at inference. -The obvious question is what that costs in throughput. On an NVIDIA GB10 running Qwen3.6-27B in NVFP4, greedy, closed loop, against vLLM in its production graphed configuration rather than `--enforce-eager`: +The question is what that does to throughput. On an NVIDIA GB10 running Qwen3.6-27B in NVFP4, greedy, closed loop, against vLLM in its production graphed configuration rather than `--enforce-eager`:
@@ -31,13 +31,15 @@ The obvious question is what that costs in throughput. On an NVIDIA GB10 running
-We are ahead at all six points, and five of those six are ties. Our run-to-run noise band is 0.5%, and concurrency 2 through 32 land between 0.7% and 1.7%, so the honest reading is that only the single-stream case (4.5%) is clearly outside noise. Output is token-for-token identical to vLLM at every point on that curve. Peak host memory is 24.88 GiB against 28.18 GiB. +Those are ties. Our run-to-run noise band is 0.5%, and concurrency 2 through 32 land between 0.7% and 1.7%, so those five points sit inside the noise or close enough to it not to matter. Only the single-stream case, at 4.5%, is clearly outside. Output is token-for-token identical to vLLM at every point on the curve, and peak host memory is 24.88 GiB against 28.18 GiB. -A tie against a mature CUDA stack is a good result for a 66 MiB binary, and it means the footprint saving is not paid for in throughput. Against llama.cpp on CPU from the same GGUF file, prefill runs 1.18x faster (223.8 against 177.3 tok/s), decode is a tie inside llama.cpp's own spread, and the tokens are byte-identical to its greedy decode. Against MLX-LM on an Apple M4, prefill time to first token is 1.5% ahead and warm total throughput is 97.6% of MLX-LM, a real 2.4% gap that sits entirely in decode. +The install drops from 9.1 GiB to 66 MiB and the throughput stays where it was, which is what we were after. -## Sometimes the port is simply faster +Against llama.cpp on CPU from the same GGUF file, prefill runs 1.18x faster (223.8 against 177.3 tok/s), decode is a tie inside llama.cpp's own spread, and the tokens match its greedy decode exactly. Against MLX-LM on an Apple M4, prefill time to first token is 1.5% ahead and warm total throughput is 97.6% of MLX-LM, a real 2.4% gap that sits entirely in decode. -[depth-anything.cpp](https://github.com/mudler/depth-anything.cpp) is a port of ByteDance's Depth Anything 3, which gives you metric depth in metres from one ordinary photo, plus per-pixel confidence, camera intrinsics and extrinsics, and a back-projected point cloud. On CPU it is faster than PyTorch running the same model. +## depth-anything.cpp is faster on CPU + +[depth-anything.cpp](https://github.com/mudler/depth-anything.cpp) is a port of ByteDance's Depth Anything 3, which gives you metric depth in metres from one ordinary photo, plus per-pixel confidence, camera intrinsics and extrinsics, and a back-projected point cloud. On CPU it runs faster than PyTorch on the same model.
@@ -49,40 +51,42 @@ A tie against a mature CUDA stack is a good result for a 66 MiB binary, and it m
-Same model, 1.31x the speed, 27% of the memory, and a load that finishes in 40 ms instead of 749 ms, on a Ryzen 9 9950X3D at 504x336 with 16 threads. The quantized q4_k build is a 99 MB file and stays near-lossless. Output correlates 1.0 with the reference forward pass, component by component, across 37 parity tests. +That is on a Ryzen 9 9950X3D at 504x336 with 16 threads. The C++ build runs the same model 1.31x faster, uses 363 MB of RAM against 1328 MB, and loads in 40 ms instead of 749 ms. The quantized q4_k build is a 99 MB file and stays near-lossless. Output correlates 1.0 with the reference forward pass across 37 parity tests. -The reason it is faster has nothing to do with writing better matmul kernels than PyTorch. Two positional embeddings, the DPT head's UV embedding and the backbone's bicubic position embedding, were being recomputed on every forward pass with single-threaded scalar sin, cos and bicubic loops, even though they depend only on the input geometry and are identical every call. Caching them removed about 95 ms of host-side overhead per forward, which is most of the gap. PyTorch builds the same embeddings with vectorized operations and never paid that cost. +We did not write a better matmul kernel than PyTorch. Two positional embeddings, the DPT head's UV embedding and the backbone's bicubic position embedding, were being recomputed on every forward pass with single-threaded scalar sin, cos and bicubic loops, even though they depend only on the input geometry and are identical every call. Caching them removed about 95 ms of host-side overhead per forward, which is most of the gap. PyTorch builds the same embeddings with vectorized operations and never had that overhead to begin with. -That is the general shape of these wins. The heavy GEMMs are close to a wash, because everyone is calling into the same class of BLAS kernel. The difference sits in host-side work that a Python reference implementation never bothered to optimize, and in not loading an interpreter and a framework to do inference. On GPU the picture flips back to parity: with the ggml CUDA backend and flash attention on a GB10, depth-anything.cpp ties PyTorch's tuned cuDNN at 47.3 ms per forward, and wins only the cold start, loading 1.75x to 2.9x faster. +The heavy GEMMs are close to a wash, because everyone is calling into the same class of BLAS kernel. What is left is host-side work that a Python reference implementation never bothered to optimize, plus not loading an interpreter and a framework to do inference. On GPU it goes back to parity: with the ggml CUDA backend and flash attention on a GB10, depth-anything.cpp ties PyTorch's tuned cuDNN at 47.3 ms per forward, and wins only the cold start, loading 1.75x to 2.9x faster. -## Parity is the gate, speed is the follow-up +## The two where we are slower -[face-detect.cpp](https://github.com/mudler/face-detect.cpp) and [voice-detect.cpp](https://github.com/localai-org/voice-detect.cpp) replaced LocalAI's Python `insightface` and `speaker-recognition` backends. Both are the case where we do not claim a CPU speed win, and both shipped anyway. +[face-detect.cpp](https://github.com/mudler/face-detect.cpp) and [voice-detect.cpp](https://github.com/localai-org/voice-detect.cpp) replaced LocalAI's Python `insightface` and `speaker-recognition` backends. Neither of them is faster than what it replaced on CPU, and we shipped them anyway. -face-detect.cpp runs the whole insightface buffalo chain, so SCRFD detection, five-landmark similarity-transform alignment to 112x112, and the ArcFace embedding, out of one self-contained GGUF with no Python and no onnxruntime. Detector boxes and landmarks match insightface to within 1 pixel, and the recognition embedding matches to cosine 1.000000, held at any thread count. On CPU it is slower than onnxruntime: SCRFD detect runs at about 0.83x at one thread and 0.69x at eight, ArcFace embed at about 0.61x and 0.84x. onnxruntime's MLAS convolution kernels sit at the FMA-port peak, and a custom AVX2 Winograd path narrowed the gap without closing it. On GPU, routing the same convolutions through cuDNN takes SCRFD from 14.8 ms to 6.4 ms and lands at torch-cuDNN parity. +face-detect.cpp runs the whole insightface buffalo chain, so SCRFD detection, five-landmark similarity-transform alignment to 112x112, and the ArcFace embedding, out of one self-contained GGUF with no Python and no onnxruntime. Detector boxes and landmarks match insightface to within 1 pixel, and the recognition embedding matches to cosine 1.000000 at any thread count. On CPU it is slower than onnxruntime: SCRFD detect runs at about 0.83x at one thread and 0.69x at eight, ArcFace embed at about 0.61x and 0.84x. onnxruntime's MLAS convolution kernels sit at the FMA-port peak, and a custom AVX2 Winograd path narrowed the gap without closing it. On GPU, routing the same convolutions through cuDNN takes SCRFD from 14.8 ms to 6.4 ms, which lands at torch-cuDNN parity. -voice-detect.cpp is the same story with a memory result attached. A WeSpeaker verification peaks at about 62 MB in our binary against about 334 MB for the CPU-only Python, torch and onnxruntime path, roughly 5.4x lower, with an identical verdict and embedding cosine 1.000000. End to end on CPU the two land within 10 to 15% of each other, trading the lead by model and thread count, and on GPU the conv encoders match the reference. +voice-detect.cpp has a memory result instead. A WeSpeaker verification peaks at about 62 MB in our binary against about 334 MB for the CPU-only Python, torch and onnxruntime path, roughly 5.4x lower, with an identical verdict and embedding cosine 1.000000. End to end on CPU the two land within 10 to 15% of each other, trading the lead by model and thread count, and on GPU the conv encoders match the reference. -For a biometric pipeline, matching the reference exactly matters more than being faster than it. An embedding that differs in the fourth decimal place changes verification decisions at a threshold, and every enrolled template in a deployment would have to be recomputed. Parity is what makes the replacement a drop-in rather than a migration. +For a biometric pipeline we would rather have the exact match than the speed. An embedding that differs in the fourth decimal place changes verification decisions at a threshold, and every enrolled template in a deployment would have to be recomputed. Matching insightface exactly is what lets somebody swap the backend out without re-enrolling their users. -## The method +## How we do it -Every port follows the same sequence, and the order is the important part. +Every port follows the same four steps. Convert the weights first, into one GGUF with the tokenizer, the vocabulary and any auxiliary model embedded, so that deploying the model is copying a file. -Port the graph second, and gate it component by component against reference tensors dumped from the original implementation. depth-anything.cpp has 37 ctest cases covering preprocessing, backbone, attention, the DPT head, depth, pose, the ray head, the ray to pose solver and the exporters. parakeet.cpp gates on transcript agreement with NeMo at WER 0. face-detect.cpp gates on box and landmark distance in pixels and embedding cosine. A port that is fast and slightly wrong is worthless, and without a per-component gate you find out it is wrong months later. +Port the graph second, and check it component by component against reference tensors dumped from the original implementation. depth-anything.cpp has 37 ctest cases covering preprocessing, backbone, attention, the DPT head, depth, pose, the ray head, the ray to pose solver and the exporters. parakeet.cpp checks transcript agreement with NeMo at WER 0. face-detect.cpp checks box and landmark distance in pixels, and embedding cosine. Skip this step and you find out the port is wrong months later, from a user, on a model you had stopped thinking about. -Optimize third, with a profiler, and only after parity holds. In parakeet.cpp the decisive win was caching a prediction-network LSTM forward pass that was 97% of transducer decode time and mostly redundant. In depth-anything.cpp it was two cached positional embeddings. Neither was a kernel rewrite, and neither would have been findable without a working baseline to profile. +Optimize third, with a profiler, and only once the parity checks pass. In parakeet.cpp the win was caching a prediction-network LSTM forward pass that was 97% of transducer decode time and mostly redundant. In depth-anything.cpp it was the two positional embeddings above. Neither was a kernel rewrite, and neither would have turned up without a working baseline to profile. Expose a flat C ABI last. LocalAI dlopens the shared library through purego and calls that ABI directly, so there is no subprocess, no gRPC hop to a Python server, and no interpreter in the serving path. -## What it costs +## What it takes to maintain -Maintenance, mostly. Each engine is a repository with its own CI, its own benchmark suite, its own GGUF conversion script and its own parity baselines, and upstream keeps releasing new checkpoints that need converter work. +Each engine is its own repository with its own CI, benchmark suite, GGUF conversion script and parity baselines, and upstream keeps releasing checkpoints that need converter work. GPU kernels are the weak spot. ggml's generic CUDA convolution and attention kernels trail NVIDIA's tuned cuDNN on the conv-heavy models, which is why face-detect.cpp needs an explicit cuDNN path to reach parity, and why parakeet.cpp's GPU margin over NeMo is a median 1.25x while its CPU margin is wider. -Porting also does not scale to everything. llama.cpp, vLLM, whisper.cpp, MLX and diffusers stay wrapped, because those projects are large, fast-moving and already excellent at what they do. We write an engine when a model has no C++ implementation, when the Python dependency is heavier than the model, or when the thing we need does not exist yet. Everything else we install from somebody else. +It also does not scale to everything. llama.cpp, vLLM, whisper.cpp, MLX and diffusers stay wrapped, because those projects are large, fast-moving and already good at what they do. We write an engine when a model has no C++ implementation, when the Python dependency is heavier than the model itself, or when the thing we need does not exist yet. The rest we install like everybody else. -Every engine listed above keeps its own benchmark suite, its parity gates and its methodology in its own repository, including the runs that did not work. The full list of them is the "Backends built by us" table in the [LocalAI README](https://github.com/mudler/LocalAI#backends-built-by-us). +One thing that confuses people reading the tree for the first time: LocalAI's own core is Go, and each backend is written in whatever its model's ecosystem needs, which is why there is C++ sitting next to Python in the same repository. + +Every engine above keeps its benchmark suite, its parity checks and its methodology in its own repository, including the runs that did not work out. The full list is the "Backends built by us" table in the [LocalAI README](https://github.com/mudler/LocalAI#backends-built-by-us). diff --git a/website/content/engines/_index.md b/website/content/engines/_index.md index f2d91f487..558987c33 100644 --- a/website/content/engines/_index.md +++ b/website/content/engines/_index.md @@ -1,5 +1,5 @@ --- title: "Engines" -description: "Nineteen native C, C++ and Go engines written by the LocalAI team. No Python at inference, checked against the reference implementation in CI, and small enough to ship as one file." +description: "Eighteen native C, C++ and Go engines written by the LocalAI team. No Python at inference, checked against the reference implementation in CI, and small enough to ship as one file." extracss: ["engines.css"] --- diff --git a/website/data/engines.yaml b/website/data/engines.yaml index ca178e442..96f8f1948 100644 --- a/website/data/engines.yaml +++ b/website/data/engines.yaml @@ -1,4 +1,4 @@ -# The nineteen native engines the LocalAI team wrote, and the one quantization +# The eighteen native engines the LocalAI team wrote, and the one quantization # recipe that feeds them. This file is the single source of truth for the # /engines/ page: the layout renders whatever is here, in this order, and adds # nothing of its own. Numbers in `highlights` come from each engine's own diff --git a/website/layouts/engines/list.html b/website/layouts/engines/list.html index 462e51498..59d4d4838 100644 --- a/website/layouts/engines/list.html +++ b/website/layouts/engines/list.html @@ -10,7 +10,7 @@

Engines we build

Eighteen engines,written from scratch.

-

Most backends wrap somebody else's engine. These do not. Each one exists because the thing we needed was a multi-gigabyte Python install, or closed, or nobody had built it yet. What you get instead is a binary and a GGUF file, checked against the reference implementation in CI, running on the machine you already own.

+

Most LocalAI backends wrap somebody else's engine. These were written from scratch, each one because the thing we needed was a multi-gigabyte Python install, or closed, or nobody had built it yet. What ships instead is a binary and a GGUF file, checked against the reference implementation in CI, running on the machine you already own.

Install LocalAI → How backends work @@ -88,8 +88,8 @@

The rule we hold them to

-

A port only ships once it matches the original.

-

Every engine here is gated against the framework it replaces, on the same input, on the same machine. That means a transcript that comes out word for word identical, boxes that land on the same pixels, or a waveform inside a stated tolerance. Speed is the part we then go and win, and the numbers on this page come out of each engine's own benchmark suite, not a marketing run.

+

We do not ship a port until it matches the original.

+

Every engine here is gated against the framework it replaces, on the same input, on the same machine. That means a transcript identical to the reference, boxes that land on the same pixels, or a waveform inside a stated tolerance. Speed work comes after that, and the numbers on this page come out of each engine's own benchmark suite.

Install LocalAI → LocalAI on GitHub ↗ diff --git a/website/layouts/index.html b/website/layouts/index.html index 3c6467f0d..9cec35949 100644 --- a/website/layouts/index.html +++ b/website/layouts/index.html @@ -37,7 +37,7 @@

The runtime

-

LocalAI is the engine everything else plugs into.

+

Everything else plugs into LocalAI.

One binary with an OpenAI-compatible API in front of it. Point an existing client at it and the calls keep working, except now the model is on your machine. It also speaks the Anthropic, Ollama and ElevenLabs APIs, so most tools need a URL change and nothing else.

Underneath, a small core pulls each engine in as a separate backend, only when a model asks for it. That is why one install covers this much ground without becoming a 9 GB download.

@@ -75,7 +75,7 @@

01 / HARDWARE

Every feature ships a CPU path first.

-

Not a degraded mode that technically runs. The real one, tested in CI, on the hardware most people already have. GPUs make it faster, they are not the price of entry.

+

That path is tested in CI, on the hardware most people already have, and it is not a degraded fallback. A GPU makes it faster but is not required.

x86_64 · ARM64 · CUDA · ROCm · SYCL · Metal · Vulkan

@@ -86,7 +86,7 @@

03 / DISTRIBUTED

-

Plug in a second machine and stop there.

+

Add a second machine.

Routing, VRAM-aware placement, prefix-cache affinity and failover are the runtime's problem. You add hardware, the cluster works out what to do with it.

Smart routing · autoscaling · P2P · NATS · federation

@@ -174,7 +174,7 @@

parakeet.cpp

Twenty-seven times faster than whisper.cpp, on a CPU.

-

NVIDIA NeMo Parakeet, ported to C++ and ggml. Ten checkpoints, all of them verified at WER 0 against NeMo, which means the transcript comes out byte for byte identical while finishing first. Cache-aware streaming with end-of-utterance detection handles live audio, and the multilingual streaming model covers 40 or more locales.

+

NVIDIA NeMo Parakeet, ported to C++ and ggml. Ten checkpoints, all of them verified at WER 0 against NeMo, which means the transcript is identical to NeMo's while finishing first. Cache-aware streaming with end-of-utterance detection handles live audio, and the multilingual streaming model covers 40 or more locales.

27xvs whisper.cpp, CPU
1.40xvs NeMo, CPU median
@@ -395,7 +395,7 @@

Distributed mode with VRAM-aware routing, autoscaling, multi-user auth and per-user quotas.

MAY 2026

It sees and hears

Voice recognition, face recognition with liveness, diarization, video generation, drop-in Ollama API.

-

JUL 2026

Nineteen engines of our own

+

JUL 2026

Eighteen engines of our own

The native C and C++ ports take over the heavy Python backends, one modality at a time.