∇Continuous Function
Concept notebookChecking saved investigationReading browser-local route memory before showing a continuation.

MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism

Serving MoE turns sparse compute into a scheduling problem: routing skew can create stragglers and token-dispatch communication can bottleneck, motivating scheduling and, in systems such as MegaScale-Infer, disaggregated expert-parallel serving.

published · difficulty 4/5 · 20 min read

01

01

Intuition

Build the mental picture first so the rest of the page has something to attach to.

PredictName the object in plain language, then predict what should change.Leave with one reusable mental picture before notation appears.

MoE looks like "cheaper inference": only run a few experts per token.

In serving, the story flips. MoE becomes a systems scheduling problem:

  • tokens get routed unevenly, so some experts become bottlenecks (stragglers),
  • expert-parallel MoE layers must dispatch token activations to the selected experts and combine results; this is often all-to-all in colocated expert parallelism and becomes M2N/N2M in disaggregated layouts such as MegaScale-Infer,
  • at decode time, you also carry the KV cache burden, so you are juggling both memory and communication.

MegaScale-Infer illustrates the systems goal: keep GPUs busy despite sparsity, skew, and communication overhead. The biggest wins are not guaranteed by sparse FLOPs alone; they come from the surrounding scheduling, batching, and communication plan.

Section prompt

Copy-only prompts — each action copies a page-grounded prompt to your clipboard. Nothing is sent by this site.

02

02

Math

Translate the story into symbols, assumptions, and a derivation you can inspect.

InspectTrack the same object through the notation and check each symbol.Leave with the invariant the equations preserve.
Equation 1Inspect equation objectBytes_{comm} \approx 2\cdot T\cdot k\cdot d_{model}\cdot b,Equation 2Inspect equation objectt_{layer} \approx \max_e t_e \propto \max_e n_e.

Token dispatch communication (why layout matters)

Per MoE layer, tokens are routed to experts and then combined back. A crude but useful byte-count is:

Bytescomm≈2⋅T⋅k⋅dmodel⋅b,\mathrm{Bytes}_{\mathrm{comm}} \approx 2\cdot T\cdot k\cdot d_{\mathrm{model}}\cdot b,Bytescomm​≈2⋅T⋅k⋅dmodel​⋅b,

where:

  • TTT is tokens in the microbatch,
  • kkk is top-kkk experts per token,
  • dmodeld_{\mathrm{model}}dmodel​ is hidden width,
  • bbb is bytes per element (e.g., 2 for fp16),
  • the factor of 2 is dispatch + combine.

If communication is the bottleneck, adding more experts can make you slower even if compute drops.

Stragglers from routing skew

Let nen_ene​ be the number of tokens routed to expert eee in a microbatch. If expert compute time is proportional to token count, then layer time is dominated by:

tlayer≈max⁡e  te∝max⁡e  ne.t_{\mathrm{layer}} \approx \max_e\; t_e \propto \max_e\; n_e.tlayer​≈emax​te​∝emax​ne​.

Even if the average load is small, the maximum load can be much larger when routing is skewed. Tail latency kills throughput.

Disaggregation (separate pools)

One modern response is disaggregation: run attention in one pool of GPUs and experts in another, pipelining between them. This trades extra communication for higher utilization and better resource matching.

Section prompt

Copy-only prompts — each action copies a page-grounded prompt to your clipboard. Nothing is sent by this site.

03

03

Code

Keep the implementation aligned with the notation so the algorithm is legible.

TraceMatch variables to symbols before reading the implementation.Leave with a runnable witness for the math.
Code witness 1import numpy as np rng = np.random.default_rng(0) E = 16 # experts T = 4096 # tokens in a mic...python
import numpy as np

rng = np.random.default_rng(0)
E = 16        # experts
T = 4096      # tokens in a microbatch

# A skewed routing distribution (one "popular" expert)
p = np.ones(E) * 0.9
p[0] = 6.0
p = p / p.sum()

choices = rng.choice(E, size=T, p=p)  # top-1 for simplicity
counts = np.bincount(choices, minlength=E)

print("avg tokens/expert:", round(float(counts.mean()), 2))
print("max tokens/expert:", int(counts.max()))
print("straggler factor max/mean:", round(float(counts.max() / counts.mean()), 2))
print("top-5 expert loads:", np.sort(counts)[-5:][::-1].tolist())
Section prompt

Copy-only prompts — each action copies a page-grounded prompt to your clipboard. Nothing is sent by this site.

04

04

Interactive Demo

Use direct manipulation to connect the explanation to a moving system.

ManipulateChange one control and predict the visible response before reveal.Leave with the observed invariant or a repaired model.

Live Concept Demo

Explore MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism

The stage is code-native and interactive. Use it to test the explanation against the mechanism.

difficulty 4/5graduatecode-aligned
Demo inquiry checkpoint

Manipulate one control and predict the visible change.

01Choose lensTrace a quantity
02ObserveDemo state pending
03GroundName the equation, invariant, or control that explains it.
04CarryNext: Speculative Decoding: Lossless Multi-Token Generation

Choose what to inspect in MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism. This shared fallback is an observation guide, not evidence of learning.

Loading interactive demo...

The demo below asks you to predict the serving bottleneck before revealing expert loads and token-dispatch communication bytes. The key invariant is that sparse activated compute does not guarantee low latency: routing skew and dispatch/combine traffic can bottleneck the layer.

Section prompt

Copy-only prompts — each action copies a page-grounded prompt to your clipboard. Nothing is sent by this site.

4/4 sections ready

Concept: MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism

What is the smallest example that makes MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism click without losing the math?

BeforeScaled Dot-Product Attention & Transformer LayersNow4/4 sections readyTryManipulate one control and predict the visible change.NextSpeculative Decoding: Lossless Multi-Token Generation
Object contextLLM Systems
StateLLM Systems
3 sources attachedChecking local snapshot
concept:llm-systems/moe-serving
Math2 equation objectsBytes_{comm} \approx 2\cdot T\cdot k\cdot d_{model}\cdot b,CodeMoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated P...rng = np.random.default_rng(0)DemoMoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated P...Manipulate one control and predict the visible change.
ConceptLearner lens

MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism

What is the smallest example that makes MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism click without losing the math?

Mode questionCan I say the mechanism back in one sentence before I reveal anything?

Start with the prediction checkpoint, then compare the reveal to the mental model.

Take this move

Study modes

Keep the object fixed; change the lens.
01PredictCommit before reveal.02ExplainName the invariant.03Ground3 source links nearby.04ExperimentChange one knob.05CarrySave the next question.

Route back through the notebook

Carry the same object through intuition, math, code, and demo.

4/4 sections ready
Carry inScaled Dot-Product Attention & Transformer Layers

Bring the mental model from Scaled Dot-Product Attention & Transformer Layers; this page will reuse it instead of restarting from zero.

Work hereMoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism

Serving MoE turns sparse compute into a scheduling problem: routing skew can create stragglers and token-dispatch communication can bottleneck, motivating scheduling and, in systems such as MegaScale-Infer, disaggregated expert-parallel serving.

Carry outSpeculative Decoding: Lossless Multi-Token Generation

The next edge should feel earned: use the demo prediction here before following Speculative Decoding: Lossless Multi-Token Generation.

Test the linkManipulate one control and predict the visible change.Then continue to Speculative Decoding: Lossless Multi-Token Generation

After The First Pass

Turn the concept into an inspected object.

The lower panels are one second act: keep the object fixed, inspect it visually, check source boundaries, practice transfer, then attach the research question.
ConceptMoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated ParallelismLLM Systems
01MechanismInspect motionSee how the idea changes form before adding more explanation.02VisualChoose a cueBind one visible mark to one mathematical role.03SourcesCheck boundary3 source links attached to the object.04ClaimAudit evidenceTreat the page claim as something to verify.05PracticeAttempt transferWrite before feedback so the model has something to repair.06RoomAsk groundedAttach the question to the same exact object.

Mechanism Storyboard

See the idea move before the page explains it

Serving MoE turns sparse compute into a scheduling problem: routing skew can create stragglers and token-dispatch communication can bottleneck, motivating scheduling and, in systems such as MegaScale-Infer, disaggregated expert-parallel serving.

Demo notes open01 / Intuition
Editorial systems illustration of sparse token routing across experts with all-to-all exchange and a straggler bottleneck.
Prediction lens

Start with the picture, metaphor, or geometric mechanism.

Commit first

Before reading further, choose the kind of change MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism should make visible.

Visual Inquiry

Make the image answer a mathematical question

Serving MoE turns sparse compute into a scheduling problem: routing skew can create stragglers and token-dispatch communication can bottleneck, motivating scheduling and, in systems such as MegaScale-Infer, disaggregated expert-parallel serving.

4/4 stages readyDemo notes connected
Prediction

Which visible object should carry the first intuition?

Commit first

Pick the cue that should make MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism easier to reason about before the page gives the answer.

Source Grounding

Canonical references for the mechanism on this page.

Object - ConceptMoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated ParallelismQuestion

What is the smallest example that makes MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism click without losing the math?

concept:llm-systems/moe-serving
Boundary

sources: shazeer-2017-sparsely-gated-moe, fedus-2021-switch-transformer, zhu-2025-megascale-infer

Check

Open the closest source note before trusting the local explanation.

Evidence

3 selected-object sources shown first; 3 references total.

Next move

Audit the claim boundary, then ask from the same selected object.

selected object source · paper · 2017Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts LayerShazeer et al.
Located CF editorial boundary

Grounds sparse expert routing and why gating turns capacity into dispatch and load-balance constraints.

Used here as

Shazeer supports sparse top-k MoE routing plus importance/load-balancing losses for uneven expert use. Fedus supports top-1 routing, expert-capacity overflow, auxiliary load balancing, an...

Caveat

Checks sparse routing, uneven expert load, and load-balancing/scheduling only. Shazeer/Fedus ground model/distributed mechanics; MegaScale-Infer supplies serving scope. Local byte-count,...

Open source
selected object source · paper · 2021Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient SparsityFedus, Zoph, and Shazeer
Located CF editorial boundary

Grounds top-1 routing, capacity factors, and communication/training-stability caveats relevant to serving.

Used here as

Shazeer supports sparse top-k MoE routing plus importance/load-balancing losses for uneven expert use. Fedus supports top-1 routing, expert-capacity overflow, auxiliary load balancing, an...

Caveat

Checks sparse routing, uneven expert load, and load-balancing/scheduling only. Shazeer/Fedus ground model/distributed mechanics; MegaScale-Infer supplies serving scope. Local byte-count,...

Open source
selected object source · paper · 2025MegaScale-Infer: Serving Mixture-of-Experts at Scale with Disaggregated Expert ParallelismZhu et al.
Located CF editorial boundary

Serving-specific evidence for MoE decoding sparsity, attention/FFN disaggregation, M2N token-routing communication, and production expert-load imbalance.

Used here as

Shazeer supports sparse top-k MoE routing plus importance/load-balancing losses for uneven expert use. Fedus supports top-1 routing, expert-capacity overflow, auxiliary load balancing, an...

Caveat

Checks sparse routing, uneven expert load, and load-balancing/scheduling only. Shazeer/Fedus ground model/distributed mechanics; MegaScale-Infer supplies serving scope. Local byte-count,...

Open source

Claim Review

Serving MoE turns sparse compute into a scheduling problem: routing skew can create stragglers and token-dispatch communication can bottleneck, motivating scheduling and, in systems such as MegaScale-Infer, disaggregated expert-parallel serving.

Object - ConceptMoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated ParallelismQuestion

What is the smallest example that makes MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism click without losing the math?

concept:llm-systems/moe-serving
Boundary

sources: shazeer-2017-sparsely-gated-moe, fedus-2021-switch-transformer, zhu-2025-megascale-infer

Check

Treat every claim as provisional until source support and a local witness agree.

Evidence

1 structured claim check on this concept.

Next move

Run the prediction or practice transfer before asking for a grounded review.

1 CF editorial source-scope review recorded

Publisher-side editorial review is not independent replication. Claims without it still need exact source-support review. 3 references and 3 local witnesses are available for inspection.

MoE inference and serving use sparse routing: a gate sends each token to a small subset of experts, so per-expert token counts can become uneven and require load-balancing, capacity buffers, or scheduling/parallelism choices.
Registered CF editorial boundary
Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts LayerGrounds sparse expert routing and why gating turns capacity into dispatch and load-balance constraints.Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient SparsityGrounds top-1 routing, capacity factors, and communication/training-stability caveats relevant to serving.MegaScale-Infer: Serving Mixture-of-Experts at Scale with Disaggregated Expert ParallelismServing-specific evidence for MoE decoding sparsity, attention/FFN disaggregation, M2N token-routing communication, and production expert-load imba...
Used here as

Shazeer supports sparse top-k MoE routing plus importance/load-balancing losses for uneven expert use. Fedus supports top-1 routing, expert-capacity overflow, auxiliary load balancing, and communication-cost...

Local witness
source-span-shazeer-2017-sparsely-gated-moesource-span-fedus-2021-switch-transformersource-span-zhu-2025-megascale-infermath-object-2code-witness-1interactive-demo
Equation 2
tlayer≈max⁡e  te∝max⁡e  ne.t_{\mathrm{layer}} \approx \max_e\; t_e \propto \max_e\; n_e.
Code witness 1import numpy as np rng = np.random.default_rng(0) E = 16 # experts T = 4096 # tokens in a mic...
Caveat

Checks sparse routing, uneven expert load, and load-balancing/scheduling only. Shazeer/Fedus ground model/distributed mechanics; MegaScale-Infer supplies serving scope. Local byte-count, straggler formula, c...

Review stateCF editorial source-scope reviewClaim metadata: source checkedPublisher-side editorial review only; not independent replication. Check caveats and exact source scope.

Shazeer 2017 supports sparse top-k MoE routing and load/importance-balancing losses for uneven expert use. Fedus 2021 supports top-1 routing, capacity overflow, auxiliary load balancing, and communication-cost tradeoffs. MegaScale-Infer 2025 supplies serving evidence for MoE decoding, token dispatch, FFN underutilization from sparsity, attention/FFN disaggregation, M2N/N2M communication in that setup, and production expert-load imbalance. Local math/code/demo remain toy witnesses.

Reviewer: codex+oracle; reviewed 2026-05-07

Practice notebook

Use the idea, then test it somewhere new

Serving MoE turns sparse compute into a scheduling problem: routing skew can create stragglers and token-dispatch communication can bottleneck, motivating scheduling and, in systems such as MegaScale-Infer, disaggregated expert-parallel serving.

AttemptNo learning claim inferred
Object - ConceptMoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated ParallelismQuestion

What is the smallest example that makes MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism click without losing the math?

concept:llm-systems/moe-serving
Boundary

sources: shazeer-2017-sparsely-gated-moe, fedus-2021-switch-transformer, zhu-2025-megascale-infer

Check

Use one state from MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism to explain what changes, why it changes, and which assumption the explanation needs.

Evidence

No learner move yet; no learning state is inferred.

Next move

Write first, use only the help you need, then try a new case without it.

Explain

Use one state from MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism to explain what changes, why it changes, and which assumption the explanation needs.

Hint 1

Reveal when your model needs a nudge.

Hint 2

Reveal when your model needs a nudge.

Hint 3

Reveal when your model needs a nudge.

Grounded object roomClose
Selected object routeAsk from this object; carry one invariant back.sources: shazeer-2017-sparsely-gated-moe, fedus-2021-switch-transformer, zhu-2025-megascale-infer
  1. ObjectConceptMoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated P...
  2. PredictBefore revealMoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated P...
  3. WitnessCompare codeMoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated P...
  4. RoomAsk groundedChecking local snapshot
ConceptMoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated ParallelismLLM Systems
Code witness comparisonMoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism code wi...rng = np.random.default_rng(0)Prediction before revealMoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism predictionManipulate one control and predict the visible change.
Grounded room questionWhat is the smallest example that makes MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism click without losing the math?Checking local snapshot

Research Room

Attach the question to an exact object

Pick the concept, equation, source, code witness, claim, misconception, or demo state before asking for help. The handoff stays grounded to that object.
Next local actionNo local draft saved yet

Open the draft below to save one note and next action in this browser.

conceptLLM Systems

MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism

Anchored question

What is the smallest example that makes MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism click without losing the math?

Source boundaryInspect source ids: shazeer-2017-sparsely-gated-moe, fedus-2021-switch-transformer, zhu-2025-megascale-inferStable content-object key attached
Role lenses for this object

These are fixed, deterministic perspectives derived from the selected object. They do not represent people, community contributions, or independent review.

Learner evidence requestAsk what would make "MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism" feel predictable rather than familiar.
Assumption

Source ids shazeer-2017-sparsely-gated-moe, fedus-2021-switch-transformer, zhu-2025-megascale-infer must support the exact object, not just the surrounding topic.

Source-checking summary

Treat this as a mechanism object: connect the definition to one equation, code witness, or demo before broadening the discussion.

Proposed experiment

Ask the learner to perturb one representation, then check whether the same invariant survives in math, code, and demo.

Next action

The learner can state the mechanism in their own words

Evidence4 checks
PredictionChecking carried observation
ActionReady for one action
AILearner handoff ready
Open source object
01PredictionChecking browser-local route memory
02EvidenceChecking for a carried observation
03BoundaryInspect source ids: shazeer-2017-sparsely-gated-moe, fedus-2021-switch-transformer, zhu-2025-megascale-infer
04Next moveSave one next action
Local action draftNo local draft saved yetExpand only when ready to capture one local next action
Local action draft

This draft stays locally in this browser for concept:llm-systems/moe-serving.

No local draft saved.
Evidence to inspect
  • Source ids to inspect: shazeer-2017-sparsely-gated-moe, fedus-2021-switch-transformer, zhu-2025-megascale-infer
  • Definition, prerequisite, and contrast concept links
  • The equation or code witness that makes the concept operational
  • One demo state that shows the invariant instead of a slogan
What would resolve this
  • The learner can state the mechanism in their own words
  • The learner can name the prerequisite that would repair confusion
  • The learner can predict how the mechanism changes under one perturbation
Object-attached AI handoff

I am working in Continuous Function's research reading room. Object: concept - MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism Object key: concept:llm-systems/moe-serving Context: LLM Systems Anchor id: concept/concept-notebook/llm-systems/moe-serving Open question: What is the smallest example that makes MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism click without losing the math? Evidence to inspect: - Source ids to inspect: shazeer-2017-sparsely-gated-moe, fedus-2021-switch-transformer, zhu-2025-megascale-infer - Definition, prerequisite, and contrast concept links - The equation or code witness that makes the concept operational - One demo state that shows the invariant instead of a slogan Deterministic role lenses for this object: - Boundary: fixed perspectives, not people, community contributions, or independent review - Source-checking summary: Treat this as a mechanism object: connect the definition to one equation, code witness, or demo before broadening the discussion. - Proposed experiment: Ask the learner to perturb one representation, then check whether the same invariant survives in math, code, and demo. - Teach/transfer move: Turn the mechanism into one sentence that predicts a neighboring concept. - Assumptions: - Source ids shazeer-2017-sparsely-gated-moe, fedus-2021-switch-transformer, zhu-2025-megascale-infer must support the exact object, not just the surrounding topic. - The stable content-object key lets local drafts, prompts, and route memory attach without changing the source page. - The concept explanation is local atlas prose until checked against its math, code, and source support. - Prerequisite gaps should become a repair route, not a reason to leave the object vague. - Role-lens requests: - Learner: ask for "Ask what would make "MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism" feel predictable rather than familiar." | assumption: Source ids shazeer-2017-sparsely-gated-moe, fedus-2021-switch-transformer, zhu-2025-megascale-infer must support the exact object, not just the surrounding topic. | next action: The learner can state the mechanism in their own words - Researcher: ask for "Source ids to inspect: shazeer-2017-sparsely-gated-moe, fedus-2021-switch-transformer, zhu-2025-megascale-infer" | assumption: The stable content-object key lets local drafts, prompts, and route memory attach without changing the source page. | next action: The learner can name the prerequisite that would repair confusion - Experimenter: ask for "Choose one variable or condition to perturb before asking for an explanation." | assumption: The concept explanation is local atlas prose until checked against its math, code, and source support. | next action: The learner can predict how the mechanism changes under one perturbation - Professor: ask for "Find the smallest transferable rule a learner could reuse without the AI." | assumption: Prerequisite gaps should become a repair route, not a reason to leave the object vague. | next action: Teach or transfer: Turn the mechanism into one sentence that predicts a neighboring concept. What would resolve this: - The learner can state the mechanism in their own words - The learner can name the prerequisite that would repair confusion - The learner can predict how the mechanism changes under one perturbation Answer as a careful research tutor: stay source-grounded, separate verified evidence from assumptions, name the relevant math objects, and end with one next action. Current deterministic role lens for this object: - Role lens: Learner - Evidence request: Ask what would make "MoE Serving & Scheduling: Token Dispatch, All-to-All, Disaggregated Parallelism" feel predictable rather than familiar. - Assumption to keep visible: Source ids shazeer-2017-sparsely-gated-moe, fedus-2021-switch-transformer, zhu-2025-megascale-infer must support the exact object, not just the surrounding topic. - Proposed experiment: Ask the learner to perturb one representation, then check whether the same invariant survives in math, code, and demo. - Next action: The learner can state the mechanism in their own words

concept/concept-notebook/llm-systems/moe-serving concept:llm-systems/moe-serving