Pith.

Distributed, Parallel, and Cluster Computing · arXiv:2610.09307 v1 · submitted 7 Oct 2026 · 77 references

vLLM-Omni Technical Report: A Unified Serving Runtime for Omni-Modality Generation

vLLM-Omni Team

The paper says

“Interaction with intelligent systems is expanding beyond text-centric chatbots and coding agents. Speech-native assistants, visual generation and editing, world-model environments, and robot action loops require models that emit text, audio, images, video, and actions.”

more of the abstract

These models differ in execution pattern: multi-stage autoregressive omni and TTS pipelines, iterative diffusion or flow-matching generators, and longer-lived world-model or robot loops that carry state across steps. As a result, serving is no longer a single text decode loop, but a heterogeneous multi-stage workflow with cross-stage transfer, streaming, and session-shaped interaction. Existing inference stacks are typically optimized for one architecture family. LLM servers deepen autoregressive scheduling and KV management, while diffusion stacks deepen denoising and parallel generation. Neither provides a shared control plane for pipelines that emit speech, pixels, or actions through separate generators, so production deployments often fall back to ad-hoc composition across disjoint runtimes. We present vLLM-Omni, a unified serving runtime for omni-modality generation. vLLM-Omni organizes each workload as a multi-stage pipeline under a single orchestrator that admits requests, advances them across stages, and demultiplexes streaming outputs. Specialized engines and stage replicas provide compute; a connector carries heavy payloads on the data plane; and session-oriented control supports long-lived duplex, world-model, and robot workloads. This report covers the architecture (stage-level KV paths, replica pools, multi-hardware platforms, and efficiency stack) and OpenAI-compatible and OpenPI APIs for omni, TTS, image/video, world-model, robot, and duplex workloads. We evaluate on the multimodal nightly CI on H100 (TTS and MiniCPM-o on H200), focused on Qwen3-Omni.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 8 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A coherent systems architecture report whose TTS and MRv2 wins are plausible but not externally sealed, because the headline tables rest on PR-pinned profiles and nightly-CI recipes with no variance data. This line is wrong

4 major objections · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 77 references say.

The paper says, at §5.1–5.3, Tables 3 and 8

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.63):

“On the measured commit, the MRv2 profile fails at startup because shared MRv2 model-state code reads a streaming-decoder attribute that only the fused Qwen3-TTS talker defines.”

Pith says: Objection 1 of 4 · major

The two headline quantitative claims—MRv2 roughly halving Qwen3-Omni mean E2EL and RTF at C=32 (Table 3) and the fused Qwen3-TTS profile achieving 1.2–2.2× audio throughput on one GPU versus the two-stage default on two GPUs (Table 8)—rest on reproducibility conditions the paper does not fully establish.

more of objection 1

Section 5.1 states that recipes are cited under tests/dfx/perf/tests/*.json and that software SHAs are those recorded by the nightly job of each table's night; Table 3 uses the PR #8223 MRv2 profile with an unmerged one-line startup guard, and Table 8 uses PR #8259 and other PR-pinned profiles. The tables report medians over three repetitions without per-run values or run-to-run variance. A pinned-commit rerun of Tables 3 and 8 with per-repetition values and variance is required before these claims are externally sealed; otherwise the reader cannot distinguish stable system behavior from nightly-CI configuration artifacts. The disclosures are honest, but the dependencies are load-bearing rather than cosmetic.

The paper says, at §5.1, §5.2, Tables 3–5

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.70):

“During two separate runs, its thinker exited in the thirdC=32 repetition and failed 32 of 128 requests, so the table reports the default deployment.”

Pith says: Objection 2 of 4 · major

The statement in §5.1 that all reported cells complete successfully needs to be reconciled with the failures disclosed elsewhere in the same section. The MRv2 text-only route failed 32 of 128 requests in one run and is therefore not reported (§5.2), and the MRv2 profile required a one-line startup guard on the measured commit.

more of objection 2

Because the reported MRv2 speech gains in Tables 3 and 4 are central to the paper, the manuscript should state explicitly whether any reported cell experienced failed or retried requests and, if so, how those failures were handled. The current wording leaves open the possibility that the reported MRv2 numbers come from runs whose reliability is not representative.

The paper says, at §5.3, Table 8

The review did not quote this passage.

Pith says: Objection 3 of 4 · major

The Qwen3-TTS comparison is between a fused single-GPU deployment and a two-stage two-GPU default. This is a legitimate and interesting comparison, but as reported it conflates fusion with resource count and with the specific MRv2 configuration common to both arms.

more of objection 3

The 1.2–2.2× ratio is stated as audio throughput without a per-GPU or cost-normalized value, and no quality metric is reported for the Qwen3-TTS rows (Whisper WER is given for CosyVoice3, Higgs Audio v3, Fish Speech S2 Pro, and MOSS-TTS-Local but not for Qwen3-TTS). The paper should report resource-normalized throughput and a quality check for Qwen3-TTS so the headline claim is not read as a pure architecture effect.

The paper says, at §5 and §6

The review did not quote this passage.

Pith says: Objection 4 of 4 · major

The evaluation does not compare vLLM-Omni against the closest alternative runtimes named in Section 6: SGLang-Omni [22] and M* [23]. The default-versus-optimized tables are self-comparisons in which both arms are configurations selected by the authors; this is reasonable for demonstrating internal profile gains, but it does not support the broader claim that a unified orchestrator is preferable to ad-hoc composition or to existing stage-oriented runtimes.

more of objection 4

A benchmark against at least one external runtime on a common pipeline and GPU configuration would substantially strengthen the central claim.

6 minor objections

The paper says, at §5.1, Tables 3 and 14

Pith says: minor objection 1 of 6

The table titles say medians over three repetitions while the columns are labelled mean; the paper should clarify whether each cell is the median of per-run means, the mean of per-run means, or the median of all individual measurements. The distinction matters for the headline ratios.

The paper says, at §5.3

Pith says: minor objection 2 of 6

The prose notes that the Base rows of Table 8 assume a server that has already processed the reference clips and that a freshly started server has about 131 ms median TTFP at C=1; this warm-cache condition should be stated in the table caption or as a footnote so the table is not read as cold-start behavior.

show 4 more

The paper says, at §2.2, Figure 3

Pith says: minor objection 3 of 6

The control-flow diagram branches on async_chunk and batch without defining them in the caption; adding one sentence describing the two transfer modes would make the figure self-contained.

The paper says, at §5.4, Table 9

Pith says: minor objection 4 of 6

The step-execution rows at C=2, 4, and 8 show per-request latency increasing while req/s saturates between roughly 0.67 and 0.82; a sentence explaining the saturation would prevent readers from interpreting it as linear load scaling.

The paper says, at §5.6, Table 14

Pith says: minor objection 5 of 6

The duplex row counts sessions rather than requests, as the text explains; renaming that column to req/s (sessions) in the table would avoid confusion.

The paper says, at References [22], [23], [74]

Pith says: minor objection 6 of 6

The closest-alternative references are project documentation or preprints rather than stable published artifacts; the bibliography should indicate versions or release dates so the reader can locate the exact systems being compared.

What would settle it

At commit ee8fdab1d (used for Tables 3 and 5), run the Fixed-Len-2500/900 recipe and the qwen3_tts_fused_single_gpu.yaml recipe named from tests/dfx/perf/tests/*.json on H200 with the stated warm protocol and three repetitions, applying only the documented one-line MRv2 guard; require median E2EL, RTF, TTFP, and audio-s/s to reproduce Tables 3 and 8 within a pre-specified tolerance, and publish per-run logs. If the MRv2 Qwen3-Omni profile does not start at that commit without the guard, the published numbers describe a patched fork rather than the released runtime.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 8 Oct 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic agrees 1 min 33 s · 36,875 tokens read, 12,251 written
  3. Editor wrote the verdict above 26 s · 37,908 tokens read, 1,925 written
  4. Referee major revision: 4 major objections, 6 minor 1 min 19 s · 37,360 tokens read, 6,144 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

I've read vLLM-Omni. What you should know: the architecture is described carefully and the evaluation is unusually honest, but the headline numbers are not yet reproducible from the paper alone. I would still send it to peer review, with a requirement that the authors ship pinned recipes and raw per-run data.

The genuinely new material is the orchestration layer: an explicit orchestrator that owns request admission and advancement, a final-output stage set for multi-modality completion, async-chunk pre-warming for low first-packet latency, per-stage replica pools with sticky affinity, typed intermediate payloads over a connector that deliberately stays off the control plane, and session control for duplex and world-model workloads. The prior vLLM-Omni stage-graph work is extended rather than replaced, and SGLang-Omni and M* are credited as parallel designs, so the high-level idea is not new. But the specific mechanisms are concrete and mostly well explained. Sections 2 through 4 read like a system that actually exists, and the portability and efficiency sections stay appropriately scoped.

The evaluation is also better than most system reports. The tables give hardware, commits, recipe paths, and concurrency values. The authors disclose that the MRv2 Qwen3-Omni profile needed a one-line startup guard, that one text-only run failed 32 of 128 requests, and that MiniCPM-o turn mode did not stop reliably. That level of candor matters, and it makes me trust the architectural claims.

The soft spots are real and concentrated in the quantitative results. Tables 3 and 8 are medians over three runs with no per-run values and no variance. The recipes live under tests/dfx/perf/tests/*.json, and the software SHAs are whatever the nightly job recorded, with some rows week-averaged and others freeze-day. Table 3 uses an unmerged PR profile plus a one-line fix; Table 8 uses PR-pinned profiles. If any of those recipes or profiles cannot be rerun at the recorded commits, the central gains cannot be verified. The self-comparison against a default deployment is not itself a flaw, but without a pinned one-command rerun or raw logs, the ratios are not sealed. There is also no direct comparison against SGLang-Omni or M*, which are the closest stage-based alternatives. That is a gap, not a fatal one.

Who gets value: systems researchers and production inference engineers deciding whether to consolidate multimodal serving onto one runtime. The paper gives them a clear architecture and an honest status report, even though the empirical claims need follow-up.

Recommendation: accept for peer review with major revision. Ask for commit-pinned benchmark recipes, raw per-run logs, variance reporting, and at least one direct comparison with an alternative stage-based runtime before the measured gains are treated as established.

The strongest objection

Both headline quantitative claims are load-bearing: MRv2 roughly halves mean E2EL and RTF at C=32 (Table 3), and fused Qwen3-TTS raises audio throughput on one GPU against the two-stage default on two GPUs (Table 8). These tables are not produced by a fixed, self-contained benchmark. Section 5.1 cites recipes under tests/dfx/perf/tests/*.json, states that software SHAs are 'those recorded by the nightly job of each table's night,' and mixes week-averaged nightlies with freeze-day results. Table 3 uses the PR #8223 MRv2 profile plus an unmerged one-line startup guard, and Section 5.2 also reports that the MRv2 text-only route failed 32 of 128 requests in one run. Table 8 uses PR #8259 and other PR-pinned profiles. The disclosures are honest, but the reported cells are a curated subset with three-run medians and no per-run values or variance. If any recipe JSON or profile is not runnable at the recorded commit, or if run-to-run variance is comparable to the reported ratios, the central performance claims rest on artifacts the reader cannot verify from the paper alone. The architectural description can stand on its own; the quantitative 'unified runtime gains' cannot yet.

The reader and I converge on the same bottleneck: the quantitative core of the central claim is not independently verifiable from the paper as submitted. The architecture sections are unusually explicit about what the orchestrator does and does not do, and the authors disclose failures most reports would omit, including the MRv2 startup guard, the 32/128 text-only request failures, and the MiniCPM-o turn-mode stopping problem. That candor is real evidence of good methodology, but it also shows that the headline tables were produced by PR-pinned profiles and ad hoc patches rather than by a fixed release. The reader's weakest_assumption identified the nightly-CI and self-selected baseline problem; my attack sharpens it to the two numbers the reader calls strongest. Those numbers have no variance information, no raw logs, and no per-run values, so the paper cannot currently rule out that a different nightly, a different PR revision, or a re-run with more repetitions would move the reported ratios materially. None of this undermines the architectural claim that a single orchestrator can host AR, diffusion, session, and robot stages; that claim is supported by the design description and by the honest scope statements in Sections 2.6, 3.2, and 7. It does mean the performance-based portion of the central claim should remain conditional until a pinned-commit reproduction is provided, and a direct comparison with SGLang-Omni or M* would further test external validity. I therefore keep the reader's CONDITIONAL verdict unchanged.

The referee's report

If the results hold, the report makes a useful systems contribution: it provides a coherent separation of control plane from data plane, a useful taxonomy of inter-stage payloads, and one of the few serving designs that treats TTS, diffusion, world-model, and robot-policy workloads as stages of one orchestrator rather than as disjoint runtimes. The paper's strengths include explicit hardware and software identifiers for the measured tables, PR-level identification of the optimized profiles, and unusually honest disclosure of the MRv2 startup failure, the text-only route failure, and the MiniCPM-o stopping problem. The architectural description can stand on its own; the quantitative performance claims, however, are not yet externally sealed because they rest on nightly-CI recipes and PR-pinned profiles reported as three-run medians without per-run values or variance.

Pith's reading of the paper (a machine's summary, not the authors')

Generative-model serving now spans speech-native assistants, text-to-speech, diffusion image and video generators, world models that roll out future video, and robot policies that emit actions — workloads with very different execution patterns, from multi-stage autoregressive pipelines to iterative denoising to long-lived stateful loops. The paper's claim is that all of them can be organized as multi-stage pipelines under a single orchestrator, and that vLLM-Omni is that runtime: a stage binds one model component to one execution engine and one resource budget, the orchestrator owns request admission, cross-stage advancement, and streaming progress, and a connector carries heavy payloads on a separate data plane. This matters because production deployments today compose disjoint LLM, diffusion, and realtime servers by hand, duplicating admission, lifecycle, and streaming logic; if the stage-based organization is right, one control plane carries all of these workloads. As evidence, the paper reports that a fused single-stage Qwen3-TTS profile on one H200 GPU delivers 1.2–2.2x the audio throughput of the two-stage default on two GPUs, that an on-device engine profile roughly halves Qwen3-Omni end-to-end latency at high concurrency, and that packed streaming lifts CosyVoice3 throughput 16x.

What carries the argument. The stage is the central object: one stage binds one model component — thinker, talker, codec-to-waveform, diffusion transformer, or policy — to one execution engine and one explicit resource budget, and stages exchange typed payloads, namely hidden states, codec codes, paged attention-cache blocks, and latents. The carrying mechanism of the whole design is the control/data split: an orchestrator owns request admission, cross-stage advancement, streaming progress, and modality-set completion (a request finishes only when every stage in its final-output set has drained), while OmniConnector moves heavy payloads on a separate data plane. Two mechanisms do the heavy lifting. Inter-stage chunk streaming pre-submits placeholder requests on downstream stages and streams chunks over the connector while upstream generation continues, which the paper identifies as the primary lever on time-to-first-audio-packet. Session-oriented control adds a longer-lived identity, retention policy, and admission caps on top of the streaming hooks, so full-duplex, world-model, and robot loops are an extension of the same control plane rather than a second runtime. Per-stage specialization — an autoregressive worker, a generation worker, a diffusion engine, and an AR–diffusion backend — keeps paradigm-specific scheduling inside the stage while the orchestrator treats every stage uniformly.

Core claim. The central claim is architectural: for omni-modality generation, the right unit of serving is a multi-stage pipeline under a single orchestrator, not a monolithic engine with modality plugins and not a loose composition of single-paradigm servers. The orchestrator decides which stage runs next, which replica receives the work, and when a result is client-visible versus a stage handoff; it deliberately does not schedule tokens, manage the paged attention cache, denoise diffusion steps, or copy large tensors. Compute specializes inside stages — autoregressive workers for thinker and talker stages, generation workers for codec-to-waveform, a diffusion engine with step-level scheduling and optional cache ownership, and an AR–diffusion backend for world-model rollouts that keep session state across steps — and each logical stage scales through its own replica pool with sticky affinity, so requests that stream stay pinned to the replica holding their state. Typed intermediate payloads (hidden states, codec codes, packed attention-cache blocks, latents) move over a semantic-agnostic connector, shared memory on a node and transfer-engine backends across hosts, so the control loop never becomes a tensor bus. The mechanism the paper leans on hardest is inter-stage chunk streaming, which pre-warms downstream stages and pushes partial outputs over the connector so waveform decoding overlaps codec prediction; disabling it at concurrency 32 on two H100s raises Qwen3-Omni mean end-to-end latency from about 73 to 137 seconds and pushes the real-time factor toward one. Session-oriented control then extends the same machinery to interactions that outlive one generate call: full-duplex speech, world-model rollouts, and robot observation-action loops share one session identity, retention policy, and admission cap on the same control plane. The reported numbers are the payoff: the fused Qwen3-TTS single-stage profile beats the two-stage default in every measured cell while using half the GPUs, and the MRv2 engine profile roughly halves Qwen3-Omni mean end-to-end latency and real-time factor at concurrency 16 and above.

The premise it rests on. The load-bearing premise is that the reported gains are real and attributable to the runtime: the tables cite nightly CI recipes under tests/dfx/perf/tests/*.json with per-night software SHAs, the optimised profiles are the authors' own pull-request configurations compared with the authors' own default recipes, and the text itself records a one-line startup guard for the MRv2 Qwen3-Omni profile, a failed text-only run, and an unreliable turn mode, so the headline numbers cannot yet be reproduced or independently checked from this paper alone.

If the paper is right (Pith's inferences, not the authors')

Operators can serve chat, speech, image, video, world-model, and robot workloads from one process and one API surface, eliminating the duplicated admission, lifecycle, and streaming logic that disjoint multi-runtime deployments require.

When a codec decoder is light enough to share a GPU with the talker, removing the inter-stage hop beats adding a second GPU: the fused Qwen3-TTS profile delivers 1.2–2.2x audio throughput on one H200 compared with the two-stage default on two H200s, and is faster in every measured cell.

Inter-stage chunk streaming should be the default for speech paths: with it disabled at concurrency 32, Qwen3-Omni mean end-to-end latency on two H100s rises from about 73 to 137 seconds and the real-time factor approaches one.

World-model rollouts and robot policy loops no longer need a dedicated realtime server: session identity, latent continuity, fencing metadata, and idle eviction run on the same orchestrator as one-shot generation, exposed through the same OpenAI-compatible and OpenPI surfaces.

Diffusion-specific optimisation — step-to-step denoising caches, sequence and classifier-free-guidance parallelism, CPU offload, regional compilation, and AR-to-DiT attention-cache handoff — attaches inside the diffusion stage, so hybrid AR+DiT pipelines share the same replica pools and efficiency knobs as other workloads.

The evaluation compares each optimised profile against the same system's default recipe; it never runs the head-to-head that would directly price the ad-hoc composition it argues against, such as a plain text-LLM server plus a separate TTS endpoint serving the same Qwen3-Omni traffic. How large that unmeasured gap is remains my own inference from the evaluation design, not a result the paper repor

The fused-stage idea points at the paper's own known weakness: AR-to-DiT attention-cache handoff, which the paper says has uneven coverage. Fusing or tightly chunking the autoregressive and diffusion stages the way the fused TTS profile removed the codec hop could remove that handoff entirely for hybrid image and world-model pipelines — a testable extension the paper does not pursue.

Because the tables come from nightly CI recipes with per-night software SHAs, the durable claim to verify is reproducibility: pin the cited recipes and commits, rerun the tables, and the architecture claim becomes checkable by any third party. Whether the recipes reproduce is a question the paper leaves open, not something it claims to have settled.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Interaction with intelligent systems is expanding beyond text-centric chatbots and coding agents. Speech-native assistants, visual generation and editing, world-model environments, and robot action loops require models that emit text, audio, images, video, and actions. These models differ in execution pattern: multi-stage autoregressive omni and TTS pipelines, iterative diffusion or flow-matching generators, and longer-lived world-model or robot loops that carry state across steps. As a result, serving is no longer a single text decode loop, but a heterogeneous multi-stage workflow with cross-stage transfer, streaming, and session-shaped interaction. Existing inference stacks are typically optimized for one architecture family. LLM servers deepen autoregressive scheduling and KV management, while diffusion stacks deepen denoising and parallel generation. Neither provides a shared control plane for pipelines that emit speech, pixels, or actions through separate generators, so production deployments often fall back to ad-hoc composition across disjoint runtimes. We present vLLM-Omni, a unified serving runtime for omni-modality generation. vLLM-Omni organizes each workload as a multi-stage pipeline under a single orchestrator that admits requests, advances them across stages, and demultiplexes streaming outputs. Specialized engines and stage replicas provide compute; a connector carries heavy payloads on the data plane; and session-oriented control supports long-lived duplex, world-model, and robot workloads. This report covers the architecture (stage-level KV paths, replica pools, multi-hardware platforms, and efficiency stack) and OpenAI-compatible and OpenPI APIs for omni, TTS, image/video, world-model, robot, and duplex workloads. We evaluate on the multimodal nightly CI on H100 (TTS and MiniCPM-o on H200), focused on Qwen3-Omni.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-08

The authors' main claim, as the classifier picked it out 0.96

“We present vLLM-Omni, a unified serving runtime for omni-modality generation. vLLM-Omni organizes each workload as a multi-stage pipeline under a single orchestrator that admits requests, advances them across stages, and demultiplexes streaming outputs.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.10 · states limitation 0.48 · human participants 0.04 · controlled comparison 0.06 · randomized assignment 0.01 · states quantitative result 0.05 · compares with prior results 0.33 · mentions code or data availability 0.08

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Interaction with intelligent systems is expanding beyond text-centric chatbots and coding agents. background 1.00 0.04 0.02 asserted 1.00
Speech-native assistants, visual generation and editing, world-model environments, and robot action loops require models that emit text, audio, images, video, and actions. background 1.00 0.05 0.03 asserted 1.00
These models differ in execution pattern: multi-stage autoregressive omni and TTS pipelines, iterative diffusion or flow-matching generators, and longer-lived world-model or robot loops that carry state across steps. background 1.00 0.08 0.03 asserted 1.00
As a result, serving is no longer a single text decode loop, but a heterogeneous multi-stage workflow with cross-stage transfer, streaming, and session-shaped interaction. background 0.82 0.10 0.11 asserted 1.00
Existing inference stacks are typically optimized for one architecture family. background 0.69 0.05 0.03 asserted 0.99
LLM servers deepen autoregressive scheduling and KV management, while diffusion stacks deepen denoising and parallel generation. background 0.95 0.06 0.03 asserted 1.00
Neither provides a shared control plane for pipelines that emit speech, pixels, or actions through separate generators, so production deployments often fall back to ad-hoc composition across disjoint runtimes. gap 0.96 0.07 0.08 asserted 0.97
We present vLLM-Omni, a unified serving runtime for omni-modality generation. vLLM-Omni organizes each workload as a multi-stage pipeline under a single orchestrator that admits requests, advances them across stages, and demultiplexes streaming outputs. contribution 0.99 0.89 0.52 asserted 1.00
Specialized engines and stage replicas provide compute; a connector carries heavy payloads on the data plane; and session-oriented control supports long-lived duplex, world-model, and robot workloads. method 0.88 0.53 0.11 asserted 1.00
This report covers the architecture (stage-level KV paths, replica pools, multi-hardware platforms, and efficiency stack) and OpenAI-compatible and OpenPI APIs for omni, TTS, image/video, world-model, robot, and duplex workloads. contribution 0.85 0.66 0.12 asserted 0.98
We evaluate on the multimodal nightly CI on H100 (TTS and MiniCPM-o on H200), focused on Qwen3-Omni. method 0.88 0.93 0.04 asserted 0.98

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.72

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-08: 1 request, 8,412 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (11)
Figure 1 of the paper
Figure 1. Scope map from application scenarios to the multi-stage model families and serving sur
Figure 2 of the paper
Figure 2. System overview of vLLM-Omni: APIs enter through AsyncOmni and its engine facade,
Figure 3 of the paper
Figure 3. Orchestrator request flow: admit, poll, then branch to client delivery or stage advancement
Figure 4 of the paper
Figure 4. Per-stage columns stack a StagePool (clients) over EngineCore replicas; Stage 1 shows
Figure 5 of the paper
Figure 5. Four inter-stage payload kinds over OmniConnector.
Figure 6 of the paper
Figure 6. Session control: lifecycle steps, three shared concerns, and modality specializations.
Figure 7 of the paper
Figure 7. vLLM-Omni as the multi-modal rollout backend in VeRL-Omni.
Figure 8 of the paper
Figure 8. Six workload families as multi-stage patterns on one runtime.
Figure 9 of the paper
Figure 9. Qwen3-Omni Fixed-Len-2500/900 on 2×H100: async-chunk versus no-async-chunk for mean TTFT, E2EL, and audio RTF. uses 832×480, 81 frames, 4 steps (plus USP2 and 1280×720/121 frames); Cosmos3-Nano follows the official demo; MiniMax-H3 uses 1344×768, 209 frames, 8 steps. MiniCPM-o 4.5 reports chal￾lenge rows at C ∈ {1, 4, 8} and one duplex Seed-TTS row at C=1. Metrics. TTFT is the delay until the first text token leaves the thinker. TTFP is the delay until the first audio packet; it is what a listener experiences as time-to-first-sound on speech outputs. TPOT is the mean interval…
Figure 10 of the paper
Figure 10. Qwen3-Omni 3-GPU replica=2 Fixed-Len sweep: mean audio TTFP and RTF versus
Figure 11 of the paper
Figure 11. Qwen3-TTS median audio TTFP and audio throughput versus concurrency on H200,

Figures are the authors', reproduced from the paper.

References (77)
  1. [22] SGLang Team. SGLang-Omni: Multi-stage serving for omni, speech, and TTS models. Project documentation, 2026 on Pith
  2. [23] Atindra Jha, Naomi Sagan, Keisuke Kamahori, Irmak Sivgin, Rohan Sanda, Steven Gao, Mark Horowitz, Luke Zettlemoyer, Olivia Hsu, Jure Leskovec, et al. M*: A modular, extensible, serving system for multimodal models, 2026 on Pith
  3. [1] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, et al. Gpt-4 technical report, 2023 on Pith
  4. [2] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, Katie Millican, et al. Gemini: A family of highly capable multimodal models, 2023. 1https://github.com/vllm-project/vllm-omni/graphs/contrib... on Pith
  5. [3] An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report, 2025 on Pith
  6. [4] Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, et al. Qwen2.5-Omni technical report, 2025 on Pith
  7. [5] Jin Xu, Zhifang Guo, Hangrui Hu, Yunfei Chu, Xiong Wang, Jinzheng He, Yuxuan Wang, Xian Shi, Ting He, Xinfa Zhu, et al. Qwen3-Omni technical report, 2025 on Pith
  8. [6] Inclusion AI, Biao Gong, Cheng Zou, Chuanyang Zheng, Chunluan Zhou, Canxiang Yan, Chunxiang Jin, Chunjie Shen, Dandan Zheng, Fudong Wang, et al. Ming-omni: A unified multimodal model for perception and generation, 2025 on Pith
  9. [7] Zhihao Du, Qian Chen, Shiliang Zhang, Kai Hu, Heng Lu, Yexin Yang, Hangrui Hu, Siqi Zheng, Yue Gu, Ziyang Ma, et al. Cosyvoice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens, 2024 on Pith
  10. [8] Zhihao Du, Yuxuan Wang, Qian Chen, Xian Shi, Xiang Lv, Tianyu Zhao, Zhifu Gao, Yexin Yang, Changfeng Gao, Hui Wang, et al. Cosyvoice 2: Scalable streaming speech synthesis with large language models, 2024 on Pith
  11. [9] Zhihao Du, Changfeng Gao, Yuxuan Wang, Fan Yu, Tianyu Zhao, Hao Wang, Xiang Lv, Hui Wang, Chongjia Ni, Xian Shi, et al. Cosyvoice 3: Towards in-the-wild speech generation via scaling-up and post-training, 2025 on Pith
  12. [10] Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, et al. Qwen-Image technical report, 2025 on Pith
  13. [11] Wan Team, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models, 2025 on Pith
  14. [12] Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, et al. Flux.1 kontext: Flow matching for in-context image generation and editing in latent space, 2025 on Pith
  15. [13] NVIDIA, Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, et al. Cosmos world foundation model platform for physical ai, 2025 on Pith
  16. [14] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P. Foster, Pannag R. Sanketi, Quan Vuong, et al. OpenVLA: An open-source vision-language-action model. InProceedings of the 8th Conference on Robot Learning (CoRL)... on Pith
  17. [15] NVIDIA, Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi "Jim" Fan, Yu Fang, Dieter Fox, Fengyuan Hu, et al. Gr00t n1: An open foundation model for generalist humanoid robots, 2025 on Pith
  18. [16] William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4195–4205, 2023 on Pith
  19. [17] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with PagedAttention. InProceedings of the 29th Symposium on Operating Systems Principles ... on Pith
  20. [18] Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie, Chuyue Sun, Jeff Huang, Cody Hao Yu, Shiyi Cao, Christos Kozyrakis, Ion Stoica, Joseph E. Gonzalez, et al. SGLang: Efficient execution of structured language model programs. InAdvances in Neural Information Processing Systems (NeurI... on Pith
show 57 more
  1. [19] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, et al. Diffusers: State-of-the-art diffusion models. GitHub repository, 2022 on Pith
  2. [20] Jiarui Fang, Jinzhe Pan, Xibo Sun, Aoyu Li, and Jiannan Wang. xdit: an inference engine for diffusion transformers (dits) with massive parallelism, 2024 on Pith
  3. [21] Peiqi Yin, Jiangyun Zhu, Han Gao, Chenguang Zheng, Yongxiang Huang, Taichang Zhou, Ruirui Yang, Weizhi Liu, Weiqing Chen, Canlin Guo, et al. vLLM-Omni: Fully disaggregated serving for any-to-any multimodal models, 2026. 31 vLLM-Omni Technical Report on Pith
  4. [24] Ruoyu Qin, Zheming Li, Weiran He, Jialei Cui, Feng Ren, Mingxing Zhang, Yongwei Wu, Weimin Zheng, and Xinran Xu. Mooncake: Trading more storage for less computation — a KVCache-centric architecture for serving LLM chatbot. In23rd USENIX Conference on File and Storage Technolog... on Pith
  5. [25] Gursimran Singh, Xinglu Wang, Yifan Hu, Timothy Yu, Linzi Xing, Wei Jiang, Zhefeng Wang, Xiaolong Bai, Yi Li, Ying Xiong, et al. Efficiently serving large multimodal models using EPD disaggregation. In Proceedings of the 42nd International Conference on Machine Learning (ICML), 2025 on Pith
  6. [26] Yongxiang Huang, Cheung Kawai, Jingan Zhou, Yingshu Chen, openYuanrong Team, and Xibin Wu. VeRL-Omni: Easy, Fast, and Stable RL Training for Diffusion and Omni-Modality Models. https://github.com/verl-project/verl-omni, 2026 on Pith
  7. [27] Junbo Cui, Bokai Xu, Chongyi Wang, Tianyu Yu, Weiyue Sun, Yingjing Xu, Tianran Wang, Zhihui He, Wenshuo Ma, Tianchi Cai, et al. Minicpm-o 4.5: Towards real-time full-duplex omni-modal interaction, 2026 on Pith
  8. [28] Rajarshi Roy, Jonathan Raiman, Sang-gil Lee, Teodor-Dumitru Ene, Robert Kirby, Sungwon Kim, Jaehyeon Kim, and Bryan Catanzaro. Personaplex: V oice and role control for full duplex conversational speech models, 2026 on Pith
  9. [29] NVIDIA. NVIDIA Nemotron voicechat. Model release, 2026 on Pith
  10. [30] Inclusion AI, Bowen Ma, Cheng Zou, ChengKun Du, Canxiang Yan, Chunxiang Jin, Chunjie Shen, Chenyu Lian, Chengxiang Fan, Dandan Zheng, et al. Ming-flash-omni: A sparse, unified architecture for multimodal perception and generation, 2026 on Pith
  11. [31] MiMo Core Team, Dong Zhang, Gang Wang, Jinlong Xue, Kai Fang, Liang Zhao, Rui Ma, Shuhuai Ren, Shuo Liu, Tao Guo, et al. Mimo-audio: Audio language models are few-shot learners, 2025 on Pith
  12. [32] Wenfu Wang, Chenxing Li, Liqiang Zhang, Yiyang Zhao, Yuxiang Zou, Hanzhao Li, Mingyu Cui, Hao Zhang, Kun Wei, Le Xu, et al. Covo-audio technical report, 2026 on Pith
  13. [33] Jaeik Kim, Woojin Kim, Jihwan Hong, Yejoon Lee, Sieun Hyeon, Mintaek Lim, Yunseok Han, Dogeun Kim, Hoeun Lee, Hyunggeun Kim, et al. Dynin-omni: Omnimodal unified large diffusion language model, 2026 on Pith
  14. [34] Hangrui Hu, Xinfa Zhu, Ting He, Dake Guo, Bin Zhang, Xiong Wang, Zhifang Guo, Ziyue Jiang, Hongkun Hao, Zishan Guo, et al. Qwen3-TTS technical report, 2026 on Pith
  15. [35] Shijia Liao, Yuxuan Wang, Songting Liu, Yifan Cheng, Ruoyi Zhang, Tianyu Li, Shidong Li, Yisheng Zheng, Xingwei Liu, Qingzheng Wang, et al. Fish audio s2 technical report, 2026 on Pith
  16. [36] Mistral-AI, Alexander H. Liu, Alexis Tacnet, Andy Ehrenberg, Andy Lo, Chen-Yo Sun, Guillaume Lample, Henry Lagarde, Jean-Malo Delignon, Jaeyoung Kim, et al. V oxtral tts, 2026 on Pith
  17. [37] Boson AI. Higgs TTS 2. Technical blog, 2025 on Pith
  18. [38] Siyi Zhou, Yiquan Zhou, Yi He, Xun Zhou, Jinchao Wang, Wei Deng, and Jingchen Shu. Indextts2: A breakthrough in emotionally expressive and duration-controlled auto-regressive zero-shot text-to-speech, 2025 on Pith
  19. [39] Yitian Gong, Botian Jiang, Yiwei Zhao, Yucheng Yuan, Kuangwei Chen, Yaozhou Jiang, Cheng Chang, Dong Hong, Mingshu Chen, Ruixiao Li, et al. Moss-tts technical report, 2026 on Pith
  20. [40] Jiayan Cui, Zhihan Yang, Naihan Li, Jiankun Tian, Xingyu Ma, Yi Zhang, Guangyu Chen, Runxuan Yang, Yuqing Cheng, Yizhi Zhou, et al. Glm-tts technical report, 2025 on Pith
  21. [41] Han Zhu, Lingxuan Ye, Wei Kang, Zengwei Yao, Liyong Guo, Fangjun Kuang, Zhifeng Han, Weiji Zhuang, Long Lin, and Daniel Povey. Omnivoice: Towards omnilingual zero-shot text-to-speech with diffusion language models, 2026. 32 vLLM-Omni Technical Report on Pith
  22. [42] Yixuan Zhou, Guoyang Zeng, Xin Liu, Xiang Li, Renjie Yu, Ziyang Wang, Runchuan Ye, Weiyue Sun, Jiancheng Gui, Kehan Li, et al. V oxcpm: Tokenizer-free tts for context-aware speech generation and true-to-life voice cloning, 2025 on Pith
  23. [43] Zach Evans, Julian D. Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Stable audio open, 2024 on Pith
  24. [44] Image Team, Huanqia Cai, Sihan Cao, Ruoyi Du, Peng Gao, Aiming Hao, Steven Hoi, Zhaohui Hou, Shijie Huang, Dengyang Jiang, et al. Z-image: An efficient image generation foundation model with single-stream diffusion transformer, 2026 on Pith
  25. [45] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023 on Pith
  26. [46] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. InInternational Conference on Machine Learning (ICML), 2024 on Pith
  27. [47] Meituan LongCat Team, Hanghang Ma, Haoxian Tan, Jiale Huang, Junqiang Wu, Jun-Yan He, Lishuai Gao, Songlin Xiao, Xiaoming Wei, Xiaoqi Ma, et al. Longcat-image technical report, 2025 on Pith
  28. [48] Guo-Hua Wang, Liangfu Cao, Tianyu Cui, Minghao Fu, Xiaohao Chen, Pengxin Zhan, Jianshan Zhao, Lan Li, Bowen Fu, Jiaqi Liu, et al. Ovis-image technical report, 2025 on Pith
  29. [49] Jiaxiang Liu, Zhida Feng, Pengyu Zou, Zhenyu Qian, Tianrui Zhu, Jun Xia, Yuehu Dong, Yanzheng Lin, Honglin Xiong, Anqi Chen, et al. Ernie-image technical report, 2026 on Pith
  30. [50] Chenyuan Wu, Pengfei Zheng, Ruiran Yan, Shitao Xiao, Xin Luo, Yueze Wang, Wanli Li, Xiyan Jiang, Yexin Liu, Junjie Zhou, et al. Omnigen2: Towards instruction-aligned multimodal generation, 2026 on Pith
  31. [51] Qi Cai, Jingwen Chen, Yang Chen, Yehao Li, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Yiheng Zhang, Fengbin Gao, Peihan Xu, et al. Hidream-i1: A high-efficient image generative foundation model with sparse diffusion transformer, 2025 on Pith
  32. [52] Krea. Krea 2 technical report. Technical report, 2026 on Pith
  33. [53] Guoxuan Chen, Chufeng Xiao, Haoran Yang, Siyue Xie, Binxiao Huang, Ming Zhang, Cheuk Him Chau, Xinyu Fu, Yingzhao Lian, Tom S. Y . Li, et al. Boogu-image-0.1: Boosting open agentic multimodal generation via understanding under a minimal budget, 2026 on Pith
  34. [54] NextStep Team, Chunrui Han, Guopeng Li, Jingwei Wu, Quan Sun, Yan Cai, Yuang Peng, Zheng Ge, Deyu Zhou, Haomiao Tang, et al. Nextstep-1: Toward autoregressive image generation with continuous tokens at scale, 2025 on Pith
  35. [55] Z.AI. Glm-image: Auto-regressive for dense-knowledge and high-fidelity image generation, 2026 on Pith
  36. [56] Tencent Hunyuan Foundation Model Team. Hunyuanimage 3.0 technical report, 2026 on Pith
  37. [57] Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, et al. Emerging properties in unified multimodal pretraining, 2025 on Pith
  38. [58] Tao Shen, Xin Wan, Taicai Chen, Rui Zhang, Junwen Pan, Dawei Lu, Fanding Lei, Zhilin Lu, Yunfei Yang, Chen Cheng, et al. Mammothmoda2: A unified ar-diffusion framework for multimodal understanding and generation, 2025 on Pith
  39. [59] Haiwen Diao, Penghao Wu, Hanming Deng, Jiahao Wang, Shihao Bai, Silei Wu, Weichen Fan, Wenjie Ye, Wenwen Tong, Xiangyu Fan, et al. Sensenova-u1: Unifying multimodal understanding and generation with neo-unify architecture, 2026 on Pith
  40. [60] Yoav HaCohen, Benny Brazowski, Nisan Chiprut, Yaki Bitterman, Andrew Kvochko, Avishai Berkowitz, Daniel Shalem, Daphna Lifschitz, Dudu Moshe, Eitan Porat, et al. Ltx-2: Efficient joint audio-visual foundation model, 2026 on Pith
  41. [61] Junsong Chen, Yuyang Zhao, Jincheng Yu, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Daquan Zhou, Huan Ling, et al. Sana-video: Efficient video generation with block linear diffusion transformer, 2025 on Pith
  42. [62] Bing Wu, Chang Zou, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Jack Peng, Jianbing Wu, Jiangfeng Xiong, Jie Jiang, et al. Hunyuanvideo 1.5 technical report, 2025. 33 vLLM-Omni Technical Report on Pith
  43. [63] Shenghai Yuan, Yuanyang Yin, Zongjian Li, Xinwei Huang, Xiao Yang, and Li Yuan. Helios: Real real-time long video generation model, 2026 on Pith
  44. [64] Shuailei Ma, Jiaqi Liao, Xinyang Wang, Jingjing Wang, Chaoran Feng, Zijing Hu, Chong Bao, Zichen Xi, Yuqi Gan, Weisen Wang, et al. Scaling mixture-of-experts video pretraining for embodied intelligence, 2026 on Pith
  45. [65] MiniMax. MiniMax H3: An open model breaking the boundaries between tasks and modalities. Technical blog, 2026 on Pith
  46. [66] Fengyi Fu, Mengqi Huang, Shaojin Wu, Yunsheng Jiang, Yufei Huo, Hao Li, Yinghang Song, Fei Ding, Jianzhu Guo, Qian He, et al. Lance: Unified multimodal modeling by multi-task synergy, 2026 on Pith
  47. [67] Seonghyeon Ye, Yunhao Ge, Kaiyuan Zheng, Shenyuan Gao, Sihyun Yu, George Kurian, Suneel Indupuru, You Liang Tan, Chuning Zhu, Jiannan Xiang, et al. World action models are zero-shot policies, 2026 on Pith
  48. [68] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.π 0: A vision-language-action flow model for general robot control, 2026 on Pith
  49. [69] Junhao Cai, Zetao Cai, Jiafei Cao, Yilun Chen, Zeyu He, Lei Jiang, Hang Li, Hengjie Li, Yang Li, Yufei Liu, et al. Internvla-a1: Unifying understanding, generation and action for robotic manipulation, 2026 on Pith
  50. [70] Gyeong-In Yu, Joo Seong Jeong, Geon-Woo Kim, Soojeong Kim, and Byung-Gon Chun. Orca: A distributed serving system for transformer-based generative models. In16th USENIX Symposium on Operating Systems Design and Implementation (OSDI 22), pages 521–538, 2022 on Pith
  51. [71] Amey Agrawal, Nitin Kedia, Ashish Panwar, Jayashree Mohan, Nipun Kwatra, Bhargav S. Gulavani, Alexey Tumanov, and Ramachandran Ramjee. Taming throughput-latency tradeoff in LLM inference with Sarathi-Serve. In18th USENIX Symposium on Operating Systems Design and Implementation... on Pith
  52. [72] Yinmin Zhong, Shengyu Liu, Junda Chen, Jianbo Hu, Yibo Zhu, Xuanzhe Liu, Xin Jin, and Hao Zhang. DistServe: Disaggregating prefill and decoding for goodput-optimized large language model serving. In 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24)... on Pith
  53. [73] Tongxuan Liu, Tao Peng, Peijun Yang, Xiaoyang Zhao, Xiusheng Lu, Weizhe Huang, Zirui Liu, Xiaoyu Chen, Zhiwei Liang, Jun Xiong, et al. xLLM technical report, 2026 on Pith
  54. [74] SGLang Team. SGLang-Diffusion: High-performance serving for diffusion and hybrid AR+DiT models. Project documentation, 2026 on Pith
  55. [75] Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. Timestep embedding tells: It’s time to cache for video diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVP... on Pith
  56. [76] DefTruth. cache-dit: A pytorch-native and flexible inference engine with hybrid cache acceleration and parallelism for dits. Open-source software, 2025 on Pith
  57. [77] Physical Intelligence. OpenPI: Open-source robot policy models and remote inference. Open-source software, 2025. 34 on Pith
Cited by 2 Pith papers

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment