Blog

How to self-host the MinerU API on RunPod

Jun 8, 2026

Last Updated: 2026-06-08

If you’re calling the official MinerU API (mineru.net/api/v4) in production, you’ve probably hit one of three walls: the daily quota of 1,000 high-priority pages, the per-file 200 MB cap, or a compliance review asking why your documents leave your infrastructure for a third-party cloud. You can run the exact same MinerU engine yourself on RunPod Serverless, keep documents in your own bucket, drop the quota, and pay roughly $0.001 per page warm. The migration is close to drop-in: swap the client, keep your create-task / poll / download loop.

The cloud API is the right way to try MinerU. Once you’re parsing real volume, self-hosting changes the economics and the data path. Here’s why, what it costs, and how to move your code over.

Why self-host the MinerU API instead of using mineru.net?

Three reasons: cost at volume, data residency, and control. The cloud API meters a daily high-priority page quota and then deprioritizes you; a self-hosted endpoint has no quota. Self-hosted, your PDFs and their parsed output never leave your RunPod worker and your own S3 bucket. And you pin the model version, pick the GPU, and set your own concurrency.

The official MinerU API caps each file at 200 MB and gives each account a daily quota of 1,000 high-priority pages (see the published limits) before jobs drop to lower priority. That ceiling is fine for evaluation and light use. It becomes a planning problem the moment you’re ingesting thousands of pages a day on a deadline.

The data path matters too. On the SaaS, every document round-trips through mineru.net. Self-hosted, the worker pulls the file from a URL you control (or you send the bytes inline), parses on your GPU, and writes the result to your bucket. Nothing transits a vendor you have to put in a data-processing agreement.

The trade-off is honest: you run the infrastructure. There’s a cold-start cost, but FlashBoot blunts it: once a host has booted the worker, it restores from a process snapshot in ~7-8 s across scale-to-zero cycles, and only a brand-new host pays the full ~110 s (vLLM init plus loading the model into VRAM). The model is baked into the image, so nothing is downloaded at request time. If your traffic is a handful of pages a month, the cloud API’s free tier is simpler; cross into steady volume and the math flips.

What does self-hosting MinerU cost vs the cloud API?

Roughly $0.001 per page warm on a 24 GB RTX 4090, plus a ~$0.03 cold-start tax the first time a host boots the worker (≈110 s of GPU billing for vLLM init and loading the baked model into VRAM). FlashBoot then snapshots that state, so the same host restarts in ~7-8 s across scale-to-zero cycles. RunPod bills per second and scales to zero, so an idle endpoint costs nothing. The cloud API is free under its daily quota, then queues or meters you; self-hosting trades that ceiling for a steady, predictable per-page rate.

The real number depends on how well you amortize that first-boot tax. A thousand pages parsed across one warm window land near $0.001/page; a single short doc on a cold host lands closer to $0.005–$0.01 because you’re paying the one-time tax against very few pages. FlashBoot shrinks that in practice, since a host pays the full tax only once. I broke the full workload-shape math down in Serverless MinerU on RunPod: honest cost math.

One cost the marketing pages skip: the first-boot cold start (~110 s on a brand-new host) is billed GPU time. FlashBoot snapshots that boot, so the same host restarts in ~7-8 s afterward. The model is baked into the image, so there’s no model-download step at boot or request time. Budget for the first boot per host, not for every request.

How do you deploy your own MinerU endpoint on RunPod?

Deploy the open-source mineru-runpod template from the RunPod Hub, or fork it and point RunPod’s GitHub build at your fork. Create a Serverless Endpoint on a 24 GB GPU (ADA_24 / RTX 4090) with FlashBoot enabled. For full_zip_url parity with the cloud API, also set four BUCKET_* env vars pointing at an S3-compatible bucket.

The fastest path is the Hub listing: one click, fill in the deploy-time form, done. The full walkthrough (fork-and-build and bring-your-own-image included) is in the deploy guide.

The one piece worth getting right up front is object storage. The compat client returns results as a full_zip_url, which means the worker uploads the output archive to a bucket and hands back a presigned URL, exactly like the SaaS. That path needs four env vars on the endpoint:

BUCKET_ENDPOINT_URL=https://<account>.r2.cloudflarestorage.com
BUCKET_NAME=mineru-outputs
BUCKET_ACCESS_KEY_ID=<key>
BUCKET_SECRET_ACCESS_KEY=<secret>

Cloudflare R2 is a good pairing here because its egress is free, so downloading your own results costs nothing. Any S3-compatible store works (R2, Backblaze B2, MinIO, AWS S3). To get started you’ll need a RunPod account; sign up here (full disclosure: that’s a referral link). Add a few dollars of credit and it covers thousands of cold starts plus millions of warm pages.

How do you migrate your code off the MinerU API?

Install the mineru-client package and swap the official requests calls for MineruApiClient. It mirrors the cloud API’s create_task / get_task surface and returns the same response dicts, so your existing poll loop keeps working. Under the hood it requests archive_format="zip", so full_zip_url comes back as a real .zip like the SaaS does.

Install it (no version pin, so it tracks the repo):

pip install "mineru-client @ git+https://github.com/sergeyshmakov/mineru-runpod"

Here’s the before and after. The official API, polling by hand:

import requests, time

H = {"Authorization": f"Bearer {MINERU_TOKEN}"}
task_id = requests.post(
    "https://mineru.net/api/v4/extract/task",
    headers=H, json={"url": pdf_url, "model_version": "vlm"},
).json()["data"]["task_id"]

while True:
    data = requests.get(
        f"https://mineru.net/api/v4/extract/task/{task_id}", headers=H
    ).json()["data"]
    if data["state"] in ("done", "failed"):
        break
    time.sleep(2)
zip_url = data["full_zip_url"]   # then download + unzip yourself

Self-hosted, against your own endpoint:

from mineru_client import MineruApiClient

client = MineruApiClient(endpoint_id="<your-endpoint-id>", api_key="<runpod-key>")

task_id = client.create_task(pdf_url, model_version="vlm")["data"]["task_id"]
done = client.wait_for_task(task_id)        # polls to a terminal state
client.download_results(done, "./out")      # full_zip_url is a real .zip; unpacked for you

Same lifecycle, same {"code": 0, "data": {...}} response shape. The parameter names map across cleanly too: model_version to the worker’s backend, language to lang, enable_formula / enable_table straight through. The full field-by-field mapping is in Migrate from the MinerU API, and the Clients page covers the native client if you’d rather use the worker’s own (richer) request shape once you’ve moved.

One auth note: the self-hosted endpoint authenticates with your RunPod API key, also via Authorization: Bearer, so even that part of your code barely changes.

What doesn’t carry over from the cloud API?

Most of it carries; four things don’t. The compat client rejects callback (a RunPod webhook delivers a different, unsigned payload than MinerU’s signed {checksum, content} callback, so it raises instead of misleading you), and it doesn’t support extra_formats (docx/html/latex), the MinerU-HTML model, or multi-range page_ranges like "2,4-6". There’s no batch endpoint (RunPod’s queue already parallelizes individual jobs across workers, so one isn’t needed). And full_zip_url requires the BUCKET_* setup above.

The full list, so there are no surprises:

Cloud API feature	Self-hosted status
`create_task` / `get_task`, state machine, `full_zip_url`	Supported (with object storage configured)
`model_version: pipeline` / `vlm`	Supported (maps to the worker backends)
`model_version: MinerU-HTML`	Not supported, raises
`extra_formats` (docx / html / latex)	Not produced by this worker, raises
`page_ranges` multi-range (`"2,4-6"`)	One contiguous range per job, raises otherwise
`callback` + `seed`	Rejected (poll with `get_task` / `wait_for_task` instead)
Batch (`/extract/task/batch`)	Not offered; submit tasks individually and raise `workers_max` — RunPod’s queue parallelizes them across workers

Where it falls down: if your pipeline leans on webhook callbacks or exports to DOCX/LaTeX, the compat client isn’t a clean swap today. There’s no batch endpoint either, but you rarely need one: submit tasks individually and let RunPod fan them across workers_max workers (the queue parallelizes them). For everything else, the create / poll / download path behaves like the SaaS.

One more practical detail. The compat client is URL-only, mirroring the cloud API’s POST /extract/task, which doesn’t accept file uploads. For small local files you don’t need to host anything: use the native MineruClient with file_b64 instead, which sends the bytes inline (fine under RunPod’s ~20 MB request cap). I parsed a 522 KB scanned Russian invoice that way and got clean Cyrillic Markdown back, no bucket round-trip involved.

FAQ

Is the MinerU API free?

The official MinerU cloud API has a free daily quota of 1,000 high-priority pages, after which jobs run at lower priority. There’s no per-call charge inside the quota. Self-hosting removes the quota entirely and replaces it with per-second GPU billing (around $0.001 per page warm).

What are the MinerU API’s limits?

Each file is capped at 200 MB, with a daily quota of 1,000 high-priority pages plus rate limits of 50 files/minute and 5,000 files/day (see the published limits). Self-hosting on RunPod removes the daily quota; the per-file practical limit becomes your GPU’s memory and your endpoint’s job timeout rather than a fixed page count.

Can I self-host MinerU without RunPod?

Yes. MinerU is Apache-2.0 open source and runs anywhere with a CUDA GPU. RunPod Serverless is the path this template targets because it scales to zero, so a bursty workload doesn’t pay for an idle GPU. On a fixed VPS or your own hardware you’d run MinerU directly and skip the serverless wrapper.

Does the self-hosted output match the cloud API’s full_zip_url?

Yes, when the endpoint has object storage configured. The compat client requests archive_format="zip", so the worker uploads a .zip to your bucket and returns a presigned full_zip_url, the same container and field the SaaS returns. download_results fetches and unpacks it for you, and autodetects .tar.gz too if you change the format.

Is self-hosting actually cheaper than the MinerU API?

At steady volume, usually yes, because you stop being throttled by the daily quota and pay only for GPU seconds used. The crossover depends on cold-start amortization: dense traffic lands near $0.001/page, sparse one-doc-per-cold-start traffic closer to $0.005–$0.01. Below a few hundred pages a month, the cloud API’s free tier is the cheaper and simpler option.

Do my documents stay private when self-hosting?

Yes. The worker fetches each file from a URL you control or from bytes you send inline, parses on your own RunPod GPU, and writes output to your own S3 bucket. No document or result passes through a third-party parsing service, which is the usual blocker in a data-residency or compliance review.

Self-hosting MinerU isn’t about beating the cloud API on accuracy. The accuracy is identical: it’s the same model. It’s about removing the quota ceiling, keeping documents in your stack, and paying per-second instead of per-tier. If that’s the wall you’ve hit, fork the template, or deploy it from the RunPod Hub and grab a RunPod account to run the first parse.

Structuring MinerU output into a clean doc tree

Jun 4, 2026

Sergei Shmakov

Last Updated: 2026-06-04

In part 1 I parsed the whole 5,039-page ECMA-376 Part 1 standard with MinerU on RunPod: 36 clause-aligned batches, about $1.15 of GPU time, out came 36 content_list.json files. That’s where most write-ups stop. Parsed is not the same as usable. A vision-language model hands you a flat stream of typed blocks with OCR quirks and no document structure. For a coding agent to answer “what does §17.9.4 isLgl (legal numbering) say?” it needs one small, faithful, addressable file, not a 200-page batch of blocks.

This post is the post-processing half: cleaning, structuring, cross-linking, and verifying the output. The whole thing is distilled into a small, document-agnostic toolkit you can run on your own MinerU output: examples/doc-structuring/ (start with the README).

End state: 5,039 pages became 9,948 Markdown files, every section addressable, ~7,900 cross-references turned into relative links, and every tag and attribute name verified against the official schema and the source PDF.

What does MinerU’s content_list.json actually give you?

A flat, page-ordered list of typed blocks: text (with an optional text_level for headings), list, table (HTML in table_body), code (in code_body, not text), equation, and image (with a VLM content description). Mixed in is page_number/header noise. No tree, no cross-reference graph, and a long tail of OCR artifacts.

That shape is fine for a quick read. It’s wrong for retrieval. Three jobs turn it into something an agent can navigate: rebuild the structure, render each block faithfully to Markdown, and verify the names didn’t get garbled on the way through the model. Everything below is generic; nothing in the toolkit knows about ECMA-376 specifically.

How do you rebuild the document structure?

The blocks arrive in reading order, so structure is just segmenting that stream by heading. A single forward walk does it: each block belongs to the section whose heading appeared most recently. Boundaries are only ever set by a real heading, so a section can never steal a neighbour’s content. You inject one callback, heading_id(block), where all your domain logic lives.

That callback is the only place document specifics enter: numbered headings, styled text_level lines, the occasional heading MinerU buried inside a code caption. The forward walk is the part that doesn’t break, because it has no notion of page or position to get wrong. The segmenter lives in segment.py.

Then the tree itself (tree.py):

a section with sub-sections becomes a folder plus a barrel file (*-0-index.md) holding its intro prose and a child index;
a leaf becomes one file. The golden rule: never split a leaf into parts.
an agent walks root → barrel → barrel → leaf, reading one small index per level instead of one giant batch.

Naming encodes the section id plus a short slug (17-4-37-tbl-table.md), so any clause is glob-findable and the camelCase tag (instrText) survives for a regex lookup.

Why is the rendering the hard part?

Because faithful Markdown is where every MinerU artifact bites. The structuring is a clean algorithm; the rendering is a pile of special cases, each one traced to a real defect on this run. Skip any of them and the content silently corrupts: examples render empty, tables lose columns, prose gets promoted to a heading.

Every fix in render.py earned its place:

Code lives in code_body (pre-fenced) and code_caption, not text. Miss this and all 3,591 XML examples render empty.
[Example: … end example] markers have to bracket the code. They routinely land before or inside the fence.
Page-split code halves (one example broken across a page) sit directly adjacent, so merge them. Genuinely separate examples have prose between them, so they’re left alone.
Mislabelled fences (txt/asp/hcl that actually hold XML) get relabelled to xml, but only when the body has namespaced tags, so a text-output example stays text.
Tables: HTML to Markdown. Inline XML examples wrapped as $<w:…>$ math get deleted by naive tag-stripping unless you protect them. Fully-empty illustration columns get dropped.
Lists: no doubled bullets (- - foo), and ordered items keep their numbers.
A long or sentence-ending text_level block is figure text or prose, not a heading. Don’t render it as ##.
$§17.4.62$ -wrapped references and \- escaped-dash bullets get normalized.

None of these are exotic. They’re just what VLM output looks like at scale, and each one quietly damages content if you skip it.

How do you cross-link a densely-referenced spec?

ECMA-376 cites itself constantly (§17.9.11, ST_Jc (§17.18.44), and on). Two moves in crosslink.py: normalize every reference to one canonical §N.N.N form so a single regex collects them all, then turn each resolvable one into a relative Markdown link to the target’s file or barrel.

The regex is §(\d+(?:\.\d+)*|[A-Z](?:\.\d+)+), which catches both numbered clauses and lettered annexes and gives you the citation graph for free. Links are relative on purpose: they’re computed section-to-section within the tree, so they stay valid wherever you mount it. No host path baked in, no rewrite on move. On this run, 7,877 of 7,959 references (98%) became working relative links.

How do you verify the parse is actually correct?

Two independent signals, both in verify.py. First, a vocabulary check: build the canonical set of element, attribute, and type names plus enum values straight from the official XSDs, then flag any name in the tree that isn’t in it but closely resembles one. Second, a source cross-check against the PDF text layer, which is the definitive one.

The vocabulary check (Vocabulary.from_xsd([...])) is fast and needs no PDF. It catches the obvious garbles: fontAlign when the schema only knows fontAlgn. But it misses the nasty case where a misread happens to spell a different real name.

That’s what the source cross-check is for. A name a file uses that is absent from that section’s own PDF page, while a near-miss correct name is present, is a confirmed garble. The PDF text layer is independent ground truth. This catches algn misread as align: align is a real element elsewhere, so it passes the vocabulary check, but on the actual page the source says algn. The check is bounded, each token is tested only against its section’s pages and deduped, so there’s no whole-document scan and no processing hole.

Confirmed garbles feed a vetted correction map (corrections.py), applied scoped to name contexts so a garble that collides with a real name is corrected only as an attribute, never as an element. Re-run the verifier and it reports zero. On this run that fixed a consistent align→algn / fontAlign→fontAlgn class across DrawingML, plus displacedByCustomXML, t12br→tl2br, subseted, and more, each confirmed against the PDF before it was applied.

Why replace OCR’d schema dumps instead of correcting them?

Because for the annexes you already have the real source, so OCR is the wrong input. The annexes are machine-generated schema listings (the full XSD and RELAX-NG for the formats). MinerU OCR’d them like everything else, producing the same garbles: CT_Placelder for CT_Placeholder, underscores read as spaces, a dangling fragment where a split cut mid-element. The real schema files exist, so swap them in.

The generic core is schema.py. Index every declaration in the official .xsd/.rnc, work out which schema file each annex dump came from (highest declaration-name overlap), and replace each parsed declaration with the authoritative one, matched by name then kind, exact → case-insensitive → fuzzy. The authoritative kind even drives the output folder and filename, so a mis-named OCR fragment self-corrects on rebuild. Result: 99.8% (5,699 of 5,710) declarations replaced. The ~11 too garbled to match confidently keep their OCR text.

How do you close the long tail of one-off OCR damage?

The residue is per-instance damage that won’t generalize into a correction map: a \@ date-switch read as $@$ , a < read as #, a dropped ), glued attribute names. Past a point, stop writing heuristics. Let agents propose fixes and gate every one of them on the source PDF.

I ran an adversarial multi-agent fan-out. The worklist was ~100 items: every §17.16.5 field clause, enum truncations found by diffing the rendered table against the authoritative XSD, and the named defects. Each item carried its PDF ground truth. About 130 agents proposed fixes, and only PDF-verified ones were accepted: 41 patches, zero that the build couldn’t apply.

They live as a per-section overlay (apply_overlay) applied last, so each find matches the on-disk text and a stale one gets reported on the next rebuild rather than vanishing silently. After all of it, verify_against_pdf reports 0 actionable garbles. The 11 it still flags were each reviewed against the PDF as genuine, distinct OOXML names (useFirstPageNumber and firstPageNumber both exist; o:cname confirmed on p4968) and recorded as benign.

What does the finished tree look like?


Source	5,039 pages, 36 MinerU batches
Output	9,948 Markdown files (4,245 leaf clauses, 356 barrels, 5,130 split schema declarations, 1 root index)
Cross-references linked	7,877 / 7,959 (98%), relative
Verification	names vs official XSD vocab (2,058 elements + 1,806 attributes) and vs the source PDF text layer

The full worked wiring is example_pipeline.py. Note how little domain code it is: an outline loader, a heading detector, a naming scheme, a reference regex, and a correction map, all driven by CLI flags with nothing hard-coded. Everything else is the library.

Where does this fall down?

Honest limits, because the toolkit isn’t magic:

The long tail is OCR, not logic, and verification closes it, not more regex. The systematic fixes get you most of the way, the authoritative-schema swap handles the annexes, and the per-instance residue is closed by the adversarial fan-out where every proposed edit is gated by the source PDF before it lands.
Verification needs a schema and a text-layer PDF. Without an authoritative vocabulary you lose the first signal; without a real text layer (a scanned PDF) you lose the second.
Structure quality equals outline quality. The tree is only as good as the section hierarchy you feed it, here the PDF bookmarks. Garbage outline, garbage tree.

FAQ

Isn’t MinerU’s Markdown output enough?

For reading, sometimes. For an addressable, agent-navigable, verified corpus, no. You need structure (the tree), a citation graph (the cross-links), and verification against ground truth. That’s the post-processing this toolkit does on top of what MinerU emits.

Why a per-section PDF cross-check instead of just trusting the schema?

Because a garble can collide with a valid name elsewhere (align is a real element), so the schema vocabulary alone passes it. The source page is the only authority on which name belongs here. Scoping the check to the section’s own pages keeps it cheap.

Do I have to use it on ECMA-376?

No, it’s document-agnostic. Supply your own Section hierarchy and a few callbacks and it runs on any MinerU output. See the README. ECMA-376 is just the worked example.

Relative or absolute cross-links?

Relative, computed within the tree. They’re identical wherever the tree is mounted, so the output ships anywhere with zero rewriting.

What block types does content_list.json contain?

text (with an optional text_level for headings), list, table (HTML in table_body), code (in code_body), equation, and image (with a VLM content description), plus page_number and header noise you filter out.

How do you handle code that MinerU split across a page?

The two halves arrive directly adjacent in the block stream, so merge adjacent code blocks. Genuinely separate examples always have prose between them, so they’re left untouched.

If this saved you time, the easiest way to say thanks is signing up for RunPod through this link. Star the repo on GitHub for updates.

Disclosure: RunPod links in this post use a referral code that credits me at no cost to you. The post would read the same without it.

Clause-aligned batching for large PDFs on MinerU + RunPod

Jun 3, 2026

Sergei Shmakov

Last Updated: 2026-06-03

ECMA-376 Part 1 (the Office Open XML File Formats — Fundamentals and Markup Language Reference) is 5,039 pages of dense, table-heavy, XML-schema-laden specification in a single 35 MB PDF. It is the document that defines .docx, .xlsx, and .pptx down to the attribute. If you want a machine-readable, clause-addressable version of it, you have to parse all 5,039 pages, and almost everything about that page count makes a naive approach fall over.

This is the story of parsing the whole thing through the mineru-runpod serverless worker. The headline result: 36 batches, 5,039 pages, 46,637 content blocks, 4,174 tables, full contiguous coverage, ~$1.15 of GPU time. The interesting part is not the total. It’s how you cut a 5,000-page document into pieces without breaking it, which it turns out is a decision about clause structure, not page numbers.

Why not just send the whole 5,000-page PDF?

The worker accepts a file_url and parses front-to-back, so technically you could send all 5,039 pages as one job. You shouldn’t, for four reasons that all get worse with size:

All-or-nothing failure. A single job that dies at page 4,800 (OOM, a transient GPU eviction, a timeout) costs you the entire run. At ~78 minutes of GPU work (more on that below), that’s an expensive coin flip.
No resumability. One job has no natural checkpoint. If it fails you start over.
The 20 MB response cap. MinerU’s output for a few hundred pages already blows past RunPod’s ~20 MB sync-response ceiling. For 5,000 pages it isn’t close: the extracted output here was 869 MB on disk.
Memory. Holding the layout model output for thousands of pages in one process is a needless VRAM/RAM risk when the work is embarrassingly sliceable.

Batching fixes all four, but only if you batch at the right boundaries.

Why cut at clause boundaries instead of every N pages?

The obvious split is “every 100 pages.” The problem: a standard isn’t a stream of interchangeable pages, it’s a tree of clauses. Clause §17.4 (Tables) might start three lines from the bottom of a page and run for 40 pages. If a batch boundary lands in the middle of it, you’ve torn a logical unit across two parse jobs, and every downstream step (clause extraction, cross-referencing, chunking for retrieval) has to stitch it back together.

So I don’t cut by page count. I cut by clause:

Build an outline from the PDF’s 4,600+ bookmarks, giving a clause → page index for the whole document.
Place batch boundaries only at clause starts, never mid-clause.
Treat the huge top-level clauses (§17 WordprocessingML, §18 SpreadsheetML, §19 PresentationML, §20/§21 DrawingML, §22 Shared MLs, and the annexes) as mandatory anchors, so a big reference section always begins a fresh batch.
Aim for ~100 pages per batch, allow up to ~200, and accept whatever the nearest clause boundary gives.

The result was 36 batches averaging 140 pages (smallest 66, largest 238). Every batch starts and ends on a clause edge, so no clause is ever split across the seam between two parse jobs.

(One calibration gotcha specific to this PDF: printed page 1 is PDF page index 9. There’s a 9-page front-matter offset you have to fold into the bookmark→page mapping or every boundary is off by nine.)

A useful consequence: because the worker slices the PDF server-side via start_page/end_page (see the API reference), you never pre-split the PDF. You upload it once and each batch job asks for its page range out of the same source file.

Did the batches actually cover the whole document?

Yes, and this is worth verifying mechanically rather than trusting. After the run, I checked each batch’s produced page span against its planned range and confirmed the batches tile the document with no gaps and no overlaps:

Metric	Value
Batches	36
Pages	5,039 (contiguous 0–5038, 0 gaps, 0 overlaps)
Content blocks	46,637
Tables	4,174
Code blocks	3,591
Pages/batch	mean 140, min 66 (b35), max 238 (b25)
Output downloaded	~465 MB (compressed tarballs)
Output on disk	869 MB (extracted)

The contiguity check is the one piece of validation I wouldn’t skip on a document this size: it’s the difference between “the run finished” and “the run is complete.”

How was the document transported in and out?

Two different transports, for two different size problems.

Input: R2 URL. At 35 MB the PDF is well over the 20 MB inline (file_b64) limit, so it can’t ride in the request body. I put it on Cloudflare R2 and passed a public URL as file_url. The worker downloads it (≤200 MB cap) and slices the requested pages itself. One upload, 36 jobs read from it.

Output: transport="s3". Per-batch output is large (the biggest batch produced a 32 MB tarball), so embedding results in the sync response was out. With transport="s3", the worker uploads each result .tar.gz back to R2 and returns a presigned URL the client downloads and extracts. The tarball carries everything: content_list.json (the flat, typed, page-indexed block list I treat as source of truth), the rendered markdown, middle.json, and a layout-overlay PDF.

The presigned URL has a 1-hour TTL, which has a real consequence for batching: you must download each batch’s result as its job finishes, not in a sweep at the end of a 78-minute run. By then the early URLs have expired.

What GPU and backend, and what did throughput look like?

Backend: vlm-auto-engine (MinerU 2.5 Pro, the MinerU2.5-Pro-2605-1.2B vision-language model) on a 24 GB AMPERE_24 (RTX A5000-class) RunPod serverless GPU. One parse per worker (MINERU_MAX_CONCURRENCY=1: vLLM’s KV cache isn’t safe to drive from concurrent parses on a 24 GB card). For how to pick a card, see Choosing a GPU.

Across the 35 timed batches, total GPU compute was 4,674.8 s (77.9 min) at an overall 1.04 pages/sec, with individual batches ranging 0.84–1.27 pp/s depending on table density. A few representative batches:

Batch	Clause	Pages	Worker time	pp/s
b00	§1 Scope (front matter)	176	145.0 s	1.21
b01	§17 WordprocessingML	100	102.8 s	0.97
b17	§18.17.7 (functions)	176	147.9 s	1.19
b25	§21.2 DrawingML – Charts	238	231.8 s	1.03
b32	Annex L Primer	102	80.2 s	1.27

Cost worked out to roughly $0.00023/page, ~$1.15 for the whole standard. Before committing to that, a 3-page smoke test (cents, ~110–130 s dominated by cold start) validated the entire pipeline end-to-end (URL fetch → parse → R2 upload → download → extract), the cheapest insurance you can buy on a big run.

The parallelism lesson: it’s RunPod-side, not in the worker

This is the part that cost the most confusion: a single batch is already parallelized inside the worker (the VLM batches many page-images through the GPU at once), but running multiple batches at once is a RunPod scaling decision, not something you trigger by submitting more jobs.

I learned this the hard way. The client submitted 3 batches concurrently, and RunPod ran exactly one while two sat in the queue. The endpoint was configured workersMax=1: one GPU worker, one batch at a time, no matter how many jobs you fire. Raising workersMax to 3 (and matching the client’s concurrency) is what actually delivered 3×: the remaining 31 batches then finished in 27.8 minutes wall-clock. The scaling guide covers how concurrency and workersMax interact.

The mental-model fix:

Inside one job: pages are parallelized on one GPU. Already maxed.
workersMax: how many separate GPUs run separate jobs at once. This is your throughput dial.

A related myth worth busting: MinerU’s pipeline logs mention a window_size=64. That is a GPU throughput batch (how many page-images stream through the model at a time to bound VRAM), not a context window. Pages are recognized independently regardless of it, so it has zero effect on content continuity across pages. Which is exactly why clause-aligned batch boundaries matter and the internal window size doesn’t: continuity is something you protect at the batch layer, not by tuning a throughput knob.

Which clauses produced the most structure?

Block and table counts track the content shape of the standard almost perfectly: the reference-material and function-catalog clauses dominate:

Batch	Clause	Blocks	Tables
b25	§21.2 DrawingML – Charts	2,829	378
b17	§18.17.7 (spreadsheet functions)	2,805	239
b26	§22 Shared MLs	2,509	—
b10	§17.17 Miscellaneous	—	238
b11	§18 SpreadsheetML	—	224

These are the dense element/attribute reference tables that make ECMA-376 what it is. They’re a good reminder to spot-check table fidelity on exactly these batches before trusting the output downstream.

The annex schema dumps look completely different

The most striking per-batch contrast is the annexes. Annex A (W3C XML Schema), Annex B (RELAX NG) and friends are long code listings, not prose with tables, and the numbers show it. Same ~150-page batch size, radically smaller output:

Batch	Annex	Tarball
b29	Annex B (RELAX NG)	1.15 MB
b30	B.3 PresentationML	1.75 MB
b27	Annex A (XML Schema)	2.06 MB
b28	A.3 PresentationML	2.23 MB

Compare that to the prose-and-table batches that ran 31–32 MB (b10, b11) for a similar page count: roughly a 15× size difference driven entirely by content type. MinerU classifies the schema listings as code, so they compress to almost nothing relative to a table-dense reference section.

How did resumability actually work?

The runner keeps a manifest.json keyed by batch, and writes each batch’s result atomically: extract into a temporary directory, then rename into place. A batch is only marked ok after its download, extraction, and rename all succeed. Two payoffs:

Pause/resume. Midway through, I paused the run to raise workersMax (you don’t want to change cluster settings while jobs are in flight). Stopping the client abandoned the in-flight jobs, but because their downloads hadn’t completed, the manifest never marked them done, so resuming re-ran them. Completed batches were skipped. No corruption, no duplicate downloads.
Crash recovery is free. The same mechanism means any crash resumes from the last completed batch.

For a 36-job run that you might interrupt, the resumable manifest is what turns “a long fragile script” into “a process you can walk away from.”

Where this falls down / what I’d change

Honest limitations:

1-hour presign expiry forces eager download. You cannot defer pulling results to the end of a long run; download each batch as it lands. My runner does this, but it’s a constraint to design around, not a free lunch.
Clause boundaries are only as good as the outline. The whole scheme leans on the PDF’s bookmark tree being accurate and complete. A document with missing or wrong bookmarks needs a fallback (TOC parsing, heading detection) before this works.
Table/code fidelity needs spot-checking. 4,174 tables and 3,591 code blocks is a lot of structure to trust blindly; the dense reference batches (b25, b17, b11) and the annex code dumps are where I’d sample-verify first.
One GPU is the ceiling. Throughput is fundamentally workersMax × per-GPU rate. There’s no in-job trick to go faster: you pay for more workers or you wait. And more workers means more cold starts, so wall-clock and cost don’t scale perfectly linearly.

What I’d change next time: drive client concurrency directly from the endpoint’s live workersMax so the two never drift, and prune the middle.json + layout PDF from batches where I only need content_list.json. They were roughly half the on-disk footprint.

FAQ

How long does it take to parse a 5,000-page PDF with MinerU?

About 78 minutes of single-GPU compute (~1 page/sec on a 24 GB RTX A5000-class card with the VLM backend), or ~28 minutes of wall-clock at 3× worker concurrency. Cost is roughly $1.15 total at ~$0.00023/page.

Why batch at clause boundaries instead of fixed page counts?

So no logical unit is split across two parse jobs. A clause can start mid-page and span dozens of pages; cutting by page count tears it in half and forces every downstream step to reassemble it. Cutting at clause starts keeps each clause whole within a batch.

How do you handle output larger than RunPod’s 20 MB response cap?

Use transport="s3": the worker uploads each result tarball to an S3-compatible bucket (Cloudflare R2 here) and returns a presigned URL you download. Per-batch output here reached 32 MB, far past the sync-response ceiling.

Does sending more concurrent jobs make a single endpoint faster?

No. Concurrency above the endpoint’s workersMax just fills the queue. Parallelism is the number of GPU workers RunPod runs, set by workersMax. Raise that to go wider.

Do I need to split the PDF before uploading?

No. Upload the full PDF once (or host it on R2 and pass file_url); each batch job requests its page range via start_page/end_page and the worker slices server-side.

How do I know the whole document was covered?

Verify mechanically: check each batch’s produced page span against its planned range and confirm the batches tile the document with zero gaps and zero overlaps. “The run finished” and “the run is complete” are not the same claim.

Where to next

The output is 36 batches of content_list.json with page-indexed, typed blocks. The next step is joining each block’s page back to the clause outline to emit a clause-addressable tree (one compact file per clause) that an agent or a retrieval index can navigate. The clause-aligned batching is what makes that join clean: every block already lives inside exactly one clause’s batch. Part 2 walks through that post-processing: cleaning the blocks, building the tree, cross-linking, and verifying the result against the source PDF.

ECMA-376 is freely available from Ecma International; it’s used here purely as a parsing benchmark. The parsed corpus is kept in a private repository for internal use, and this post shares only the parsing process and aggregate statistics, not the standard’s content.

If you want per-phase timings (fetch / parse / package) and throughput dashboards for a run like this, the worker can ship OpenTelemetry traces and metrics to any OTLP backend. See the observability guide.

If this saved you time, the easiest way to say thanks is signing up for RunPod through this link. Star the repo on GitHub for updates.

Disclosure: RunPod links in this post use a referral code that credits me at no cost to you. The post would read the same without it.

Fix RunPod's 'no resources to deploy your pod' error

Jun 3, 2026

Sergei Shmakov

Last Updated: 2026-06-03

If RunPod fails a deploy with this:

This machine does not have the resources to deploy your pod. Please try a different machine.

your Docker image is fine. This is a capacity error: RunPod’s scheduler tried to place your pod on a host that didn’t have a free GPU of the type you asked for, and bailed. It’s transient. The fix is to make RunPod try again on a different host, and how you trigger that retry depends on which RunPod product you’re using. On a Serverless endpoint wired to GitHub, push any commit to the watched branch. On a RunPod Hub template, cut a new GitHub Release. Those two triggers are not interchangeable, which is the part that trips people up.

I hit this constantly while maintaining the mineru-runpod template. The rest of this post is what the error actually means, why it’s not your fault, and the exact retry mechanic for each workflow.

What does “this machine does not have the resources to deploy your pod” mean on RunPod?

It means RunPod’s scheduler picked a physical host to run your pod, found that host couldn’t satisfy the requested GPU, RAM, or disk, and refused the placement. It’s a per-machine capacity miss, not a global outage and not a problem with your image. “Try a different machine” is literal advice: another host of the same GPU type may have room.

The message fires during the scheduling phase, before your container ever starts. That timing is the tell. A broken image fails differently: you’d see an image-pull error, a non-zero container exit, or a failed health check. This message means the scheduler never got that far. It looked at the GPU type you requested, compared it against free capacity on the candidate host, and found the fit impossible.

For most people the requested GPU is the binding constraint. Popular pools like the RTX 4090 get contended, especially in a busy region. When every host of that type in that data center is full, a fresh placement attempt fails until one frees up.

Is it a RunPod bug, or did I break something?

Neither. It’s a legitimate, expected response from RunPod’s scheduler reporting that the GPU pool you targeted had no free host at that moment. Your Dockerfile, your handler, and your config are all irrelevant to this error. GPU capacity on RunPod fluctuates minute to minute, so the same deploy that fails now often succeeds 30 seconds later on a different host.

The reason it feels like a bug is that it’s non-deterministic. The exact same config fails one minute and works the next, purely because the cluster’s free capacity moved. That’s also why retrying works: you’re not changing anything about your build, you’re just asking the scheduler to roll the dice again against a pool whose occupancy has shifted.

Where you actually meet this string matters. RunPod throws it whenever it spins up a real pod, which in this template’s world is two places: the Hub validator test pod that runs after a release, and a GPU Pod you launch directly. A Serverless worker that can’t find capacity usually surfaces it as workers stuck in a throttled or initializing state rather than this exact sentence, but the underlying cause and the recovery are the same: force a rebuild so RunPod reschedules.

How do I fix it on a RunPod Serverless endpoint?

Push any commit to the branch your endpoint watches. Per RunPod’s docs, “every git push to your specified branch results in an updated Endpoint”, so a no-op commit triggers a fresh build and redeploy. The new workers get scheduled again, almost always landing on a host with free capacity. You can also hit Rebuild in the RunPod console to do the same thing without a commit.

This is the path most people need. If you deployed your worker by connecting a GitHub repo, your endpoint redeploys on every push, so a trivial commit is the lowest-friction way to re-roll the scheduler:

git commit --allow-empty -m "chore: re-trigger RunPod build"
git push

The --allow-empty flag is the point: you don’t need a real change to force a rebuild, you just need a new commit on the watched branch. RunPod’s layer caching means the rebuild is fast after the first one, since only the layers that changed get rebuilt (and for an empty commit, none did).

If you’d rather not pollute history, the console’s manual Rebuild button is the cleaner equivalent. Either way you’re doing the same thing: asking RunPod to provision workers again, on hosts whose occupancy has moved since the last attempt.

How do I fix it when publishing a RunPod Hub template?

Cut a new GitHub Release. The Hub does not watch commits. Per RunPod’s publishing guide, “repository integration connects with GitHub repos using releases (not commits) for versioning”, so pushing to your branch does nothing on the Hub side. Only a new Release re-runs the build and the validator test pod, which is where this error shows up for template authors.

Here’s the trick that saves you: a Release is just a tag, and a tag can point at a commit you already have. You don’t need to change a single line of code to re-trigger the Hub. Tag the same HEAD you already shipped and publish a Release for it:

git tag v1.6.4          # same commit, new tag
git push origin v1.6.4
# then publish a GitHub Release for v1.6.4 in the UI or via gh

RunPod treats each tag as a distinct template version and re-runs the full pipeline: build the image, spin up the validator test pod from .runpod/tests.json, and index the listing (usually within an hour). If the previous Release failed only because the validator pod couldn’t get a GPU, the new Release gives it a fresh roll of the scheduler.

The catch worth stating: each retry adds a version to your Hub listing, even if two versions are byte-identical. That’s the cost of the release-driven model. It’s cosmetic, but if you retry five times you’ll have five versions, so don’t spin on it if the failure is actually persistent (see below).

Why is the retry different for Serverless vs the Hub?

Because the two products use different GitHub triggers. A Serverless endpoint rebuilds on every push to its watched branch, so a commit is your retry. The Hub builds only on a new Release tag, so a release is your retry. Pushing commits at a Hub listing does nothing; pushing releases at a Serverless endpoint isn’t how it watches for changes.

	Serverless endpoint	RunPod Hub template
Build trigger	Every push to the watched branch	New GitHub Release (tag)
Retry the deploy by	Empty commit, or Rebuild in console	New Release on the same commit
Who needs this	Anyone running a GitHub-connected worker	Template authors publishing to the Hub
Where the error appears	Workers stuck initializing / throttled	The validator test pod after a Release

Most readers are in the left column. You deployed a worker from a repo (or one-click from the Hub) and you’re iterating on it; a commit re-rolls it. The right column is for the smaller group of people authoring a Hub listing, where the validator test pod is the thing hitting the capacity wall. If you’re not publishing your own Hub template, you can ignore the release workflow entirely. For the full deploy walkthrough, the getting-started guide covers both paths.

What if re-triggering doesn’t fix it?

If the same GPU pool fails every time, the capacity miss is persistent, not transient, and rolling the scheduler won’t help. Switch to a higher-availability GPU pool, change region, or lower the resources you’re requesting. For the mineru-runpod Hub validator, that means editing the gpuTypeId in .runpod/tests.json to a pool that’s actually free, then cutting a new Release.

The template defaults its validator to "NVIDIA GeForce RTX 4090" because it has the best pool availability across RunPod’s regions. "NVIDIA RTX A5000" works too but tends to be scarcer. I’ve bounced the test pod’s GPU between the A40, the A5000, and the 4090 across releases, chasing whichever pool had capacity on a given day, and the 4090 wins most often.

Three levers when retries aren’t enough:

Change the GPU type to a less-contended pool. A 24 GB workload fits several pools; pick the one with capacity rather than the one you assumed.
Change the region if your endpoint or template pins one. Capacity is per data center, so a pool that’s full in one region can be wide open in another.
Reduce the request. Oversized container disk or volume sizes shrink the set of hosts that can fit your pod. Trim them if they’re padded.

For template authors, there’s also an escape hatch documented in the troubleshooting guide: if a release is urgent and the validator is the only blocker, rename .runpod/tests.json to .runpod/tests_.json so the Hub skips the test pod entirely. You lose all CI signal, so it’s a temporary unblock, not a default. For the GPU-pool math behind these choices, see Choosing a GPU.

FAQ

Is “this machine does not have the resources to deploy your pod” a RunPod outage?

No. It’s a per-host capacity miss, not a global outage. The scheduler tried one machine, found it full, and stopped. Other hosts of the same GPU type may have room, which is why a retry often succeeds within seconds even while RunPod is otherwise healthy.

Does pushing a commit fix the error on a RunPod Hub template?

No. The Hub builds only on new GitHub Releases, not commits. Pushing to your branch leaves the Hub listing untouched. You have to publish a new Release (a new tag, which can point at the same commit) to re-run the Hub build and its validator test pod.

How do I re-trigger a RunPod Serverless build without changing code?

Push an empty commit with git commit --allow-empty to the watched branch, or click Rebuild in the RunPod console. Both force a fresh build and redeploy, so workers get scheduled again on hosts whose free capacity has shifted since the last attempt.

Can I create a GitHub Release without a new commit?

Yes. A Release is a tag, and a tag can point at any existing commit. Tag your current HEAD and publish a Release for it. RunPod treats every tag as a new version and re-runs the build, so this re-triggers the Hub without any code change.

Why does the same deploy fail once and work the next time?

GPU capacity on RunPod fluctuates minute to minute. The same config hits a full host on one attempt and a free host on the next, with nothing about your image changing. That non-determinism is exactly why retrying is the first thing to try.

How do I stop hitting the capacity error repeatedly?

Stop targeting a contended pool. Switch gpuTypeId to a higher-availability GPU (RTX 4090 pools are usually the most available), change region, or reduce requested disk and volume sizes so more hosts can fit your pod.

Where to next

The error is annoying but harmless once you know it’s capacity and not your build. For a Serverless worker, a commit re-rolls it; for a Hub template, a Release does. If it persists, it’s a pool-availability problem, and the fix lives in your GPU choice, not your Dockerfile. The full set of Hub build failures (this one, the CUDA floor mismatch, and the 30-minute build timeout) is catalogued in the troubleshooting guide.

If this saved you a debugging session, star the repo on GitHub for updates, or open an issue if you hit a build failure that isn’t covered here.

How RunPod FlashBoot Actually Works (4-Request Test)

May 26, 2026

Sergei Shmakov

Last Updated: 2026-05-26

If you’re shipping vLLM or any heavy ML model on RunPod Serverless, you’ve probably looked at FlashBoot, ticked the checkbox, and then watched your cold starts still take 60-120 seconds. RunPod’s marketing says “1-second cold starts.” Their docs describe FlashBoot as “pre-loading container images.” Neither of those matches what most ML workloads actually see.

I ran four cold-start tests on a deployed RunPod endpoint serving a vLLM-backed PDF parser. The wall-clock numbers ranged from 7 seconds to 7 minutes. The point of this post is to explain why — what FlashBoot actually does at the systems level, when it kicks in, and how to set up your worker so it kicks in more often.

What does FlashBoot actually do?

FlashBoot is a CRIU-style process snapshot mechanism. When a worker scales to zero, RunPod captures the full process state (Python interpreter, CUDA VRAM, subprocess tree) into a snapshot on the host’s local storage. When the worker scales back up on the same host, RunPod restores from that snapshot. The restored process resumes mid-stride: model still in VRAM, vLLM engine subprocess still alive, IPC pipes still connected.

The key qualifier that RunPod’s docs don’t mention: snapshots are per (host, image SHA), not per endpoint. If the next scale-from-zero lands on a different host, there’s no snapshot to restore from. The worker boots fresh and pays the full warmup cost. Once.

The TL;DR for an ML workload: set up an eager warmup at worker boot, then let FlashBoot do its thing. Each new host pays the warmup tax once. Subsequent scale-from-zeroes on that same host get the snapshot restore and finish a typical request in single-digit seconds.

Why do “cold” starts sometimes take 7 seconds and sometimes 110?

Because they’re hitting different parts of the per-host model. Four consecutive requests against the same endpoint, single-page parse on each, with a deliberate scale-to-zero between every one:

Request	Wall-clock	Host	Snapshot?	What the worker did
1	456 s	A (post-rebuild)	none	Image pull + fitness checks + warmup (101 s) + parse (5.6 s)
2	7.6 s	A (same as R1)	yes	Snapshot restore + parse (4.7 s)
3	122 s	B (different host)	none	Fitness checks + warmup (101.5 s) + parse (5.6 s)
4	7.4 s	B (same as R3)	yes	Snapshot restore + parse (4.6 s)

First hit on a fresh host pays ~110 s for the warmup. Every subsequent restore on that same host is ~7-8 s. A new host, when RunPod’s scheduler picks one, starts the cycle over.

The 456 s on Request 1 included a one-time image pull (the worker image is ~27 GB; this was the first time that physical host had ever seen it). Strip that off and you get ~110 s of actual boot work, which matches Request 3 exactly.

How can you tell if a request hit a snapshot restore?

By what’s missing from the worker logs. A FlashBoot-restored worker skips its boot sequence entirely — no fitness checks, no Python import logs, no vLLM engine initialization, no model load. The first log line is Jobs in queue: 1, immediately followed by your handler’s “starting job” entry.

Compare a fresh boot to a snapshot restore for the same request shape:

Fresh boot (Request 3):

04:45:45  Running 7 fitness check(s)...
04:45:46  All fitness checks passed. (1285.99ms)
04:45:46  [mineru-warmup] starting (backend=vlm-auto-engine ...)
04:45:51  Using vllm-async-engine as the inference engine for VLM.
04:46:23  Initializing a V1 LLM engine (v0.11.2) ...
04:46:47  Model loading took 2.1601 GiB memory and 18.41 seconds
04:47:14  torch.compile takes 22.81 s in total
04:47:17  init engine (profile, create kv cache, warmup model) took 30.66 seconds
04:47:18  get vllm-async-engine predictor cost: 87.26s
04:47:28  [mineru-warmup] done in 101.5s
04:47:28  Jobs in queue: 1
04:47:28  Started.
04:47:28  "starting job" {...}
04:47:34  "done" {...elapsed_seconds: 5.58...}

Snapshot restore (Request 4):

04:51:25  Jobs in queue: 1
04:51:25  Started.
04:51:25  "starting job" {...}
04:51:26  Using vllm-async-engine ...   (instant — engine handle restored from snapshot)
04:51:30  "done" {...elapsed_seconds: 4.58...}

No boot sequence. Three timestamps. The vLLM engine subprocess PID from the previous boot is reused — same EngineCore_DP0 pid=NNN from the snapshot. If you grep your own worker logs for the gap between Jobs in queue: 1 and the previous activity, you’ll see whether RunPod did a fresh boot or a snapshot restore.

What does the FlashBoot snapshot preserve?

Everything that lived in the worker process at snapshot time, mediated by CRIU semantics:

Python interpreter state. Module imports stay loaded. Globals (job counters, contextvars, signal handlers) keep their values. The MinerU engine registry returns the same handles it returned before the snapshot.
GPU VRAM. Model weights (~2.16 GiB for our VLM), vLLM’s KV cache (~8.17 GiB on a 24 GB card), and captured CUDA graphs (~0.3 GiB) all survive. The first request after restore parses with the same allocations it had before.
The subprocess tree. vLLM runs its engine in a child process for memory isolation. That subprocess gets captured along with the parent and restored with its IPC pipes intact. The engine PID persists.
torch.compile cache. The JIT-compiled Dynamo / Inductor output stays valid across restore. No 22-second recompile.

What doesn’t survive: snapshot lifetime is limited. RunPod doesn’t publish the eviction policy, but obvious triggers include image rebuild (new SHA invalidates the snapshot), and presumably long enough idle on a busy host that the snapshot storage gets pushed out.

What broke before this worked? The asyncio gotcha

The “eager warmup at boot” idea is obvious in principle: run one throwaway parse during worker startup so the model is loaded and warm before any user request arrives. The implementation has one trap.

vLLM’s AsyncLLMEngine binds its IPC primitives (transports, queues) to the asyncio event loop that initialized it. If you call asyncio.run(warmup()) followed by runpod.serverless.start(), your warmup creates loop A, runs the parse, then tears loop A down when asyncio.run returns. Then runpod.serverless.start() creates loop B for serving. When the first user request tries to talk to the vLLM engine through loop B, the engine handle is bound to the now-dead loop A. Result:

"error_type": "EngineDeadError",
"error_message": "EngineCore encountered an issue. See stack trace (above) for the root cause."

The engine subprocess itself is still alive. It’s only the parent process’s IPC reference that’s broken.

The fix is to keep the warmup and the serve loop on the same asyncio event loop. RunPod’s runpod.serverless.start() internally calls asyncio.run(JobScaler.run()), but JobScaler (in runpod.serverless.modules.rp_scale) is constructible directly and its run() is an awaitable coroutine. So you can compose:

import asyncio
from runpod.serverless.modules import rp_ping, rp_scale
from runpod.serverless.modules.rp_fitness import run_fitness_checks

config = {"handler": handler, "concurrency_modifier": _concurrency_modifier, "rp_args": {}}

async def _bootstrap():
    await run_fitness_checks()
    await warmup_async()          # <- engine binds to THIS loop
    rp_ping.Heartbeat().start_ping()
    await rp_scale.JobScaler(config).run()   # <- and stays here

asyncio.run(_bootstrap())

Now both phases share one event loop. The engine handle stays valid across the warmup → serve transition. FlashBoot captures a snapshot of a process where the loop, the engine, and the IPC are all alive together. On restore, they come back together too.

This does reach into runpod-python’s internals (the runpod.serverless.modules.* submodules aren’t part of the documented public API). Cheap to guard against drift: a unit test that asserts JobScaler exists with the expected constructor and an awaitable run() method. If RunPod refactors, CI catches it before production does.

When does the warmup pay off and when doesn’t it?

Per host, not per endpoint. The math depends on your traffic pattern.

Scenario	Likely outcome
`workers_min ≥ 1` (always-on worker)	Worker stays on its host. Every request is on a fully warm worker (~5 s parse). No cold starts at all.
High-frequency endpoint, workers scale up and down fast	Same hosts get re-selected. Most cold starts are happy-path restores (~7 s).
Quiet endpoint, infrequent requests, long idle gaps	RunPod’s scheduler may pick a different host. Some cold starts will be on new hosts (~110 s).
First request after a rebuild	Always cold path. Every endpoint’s first request after a fresh image pays ~5-7 min (image pull) + ~110 s (warmup). One-time per worker host.
`MINERU_SKIP_WARMUP=1` (warmup off)	Every cold start is ~110-130 s. No per-host amortization. Don’t do this in production.

The case that stings is “quiet endpoint with sporadic traffic” — a few requests an hour, 10-minute idle gaps, RunPod bouncing between hosts. Without warmup, every cold start would be ~110-130 s. With warmup, you get a mix: some 7-second restores, some 110-second fresh boots. The mix tilts toward fast as the endpoint warms up across more hosts and RunPod’s scheduler starts re-selecting them.

If your traffic is sustained enough that you can pin a worker (workers_min=1), you skip the entire question. You’re paying for the GPU 24/7 but never paying a cold start. For workloads with even modest cost sensitivity, the warmup + FlashBoot path is the better trade.

What this means if you’re shipping vLLM on RunPod

Three takeaways from the live measurements:

Always set up an eager warmup at worker boot. Loading the model on first request is silently worse than it sounds — you don’t just pay 110 s once per cold start, you pay it every time a host doesn’t have a snapshot, AND you forfeit the per-host amortization that makes the second-hit-on-a-host cheap. Without warmup, FlashBoot has nothing to snapshot.
Compose warmup and the serving loop under one asyncio.run(). If you asyncio.run() the warmup separately, the engine dies at the loop boundary. The fix is straightforward but the failure mode is opaque (EngineDeadError 75 ms into the first request) — easy to misdiagnose as a vLLM bug.
Don’t market your cold start as “X seconds” without acknowledging the per-host mix. A snapshot-restore cold start is genuinely 7-8 seconds. A new-host cold start is ~110 s. Both are big improvements over the no-warmup baseline (~110-130 s per request, every request). But your users will see the mix, and a too-clean claim makes the bad days look broken.

The whole investigation was on a 24 GB A5000 / RTX 4090 class GPU running MinerU’s 1.2B VLM via vLLM 0.11.2. The numbers will shift on larger models (more VRAM to snapshot, longer model load on cold path) but the mechanism applies the same way. If your cold start dominates wall-clock latency on a serverless GPU workload, set up boot-time warmup, watch the worker logs for the snapshot pattern, and tune your workers_min accordingly.

FAQ

Does FlashBoot snapshot the vLLM engine subprocess?

Yes. The vLLM engine runs as a child process for memory isolation, and FlashBoot’s CRIU-style mechanism captures the full process tree including subprocesses. The engine’s PID persists across snapshot/restore, and its IPC pipes back to the parent stay connected.

Why does my cold start take 60-120 seconds even with FlashBoot enabled?

Most likely your model is being loaded lazily on first request rather than at worker boot. FlashBoot only snapshots state that already exists in the worker process when it scales to zero. If your model loads on first request, the snapshot captures a worker without the model, and every cold start has to load the model again. Move the model load to worker boot (before runpod.serverless.start()) and FlashBoot will start carrying the warm state forward.

What’s the difference between FlashBoot and a network volume?

A network volume is shared file storage attached to your worker (e.g., for model weights you don’t want to bake into the Docker image). FlashBoot is process-state preservation — it captures the running Python process, including data already loaded from disk into VRAM. They solve different problems and can be used together: a network volume avoids re-downloading model files on image pull; FlashBoot avoids re-loading them into VRAM on cold start.

Does FlashBoot work for non-GPU workloads?

The mechanism (process snapshot via CRIU or equivalent) doesn’t depend on GPU memory specifically. CPU-bound workloads with significant cold-start cost (heavy library imports, large in-memory indices, JIT compilation) should benefit similarly. The framing in this post happens to use a GPU workload because that’s where the cold-start tax is most painful.

How do I know if my worker is hitting a snapshot restore vs a fresh boot?

Check the worker logs in the RunPod dashboard. A fresh boot shows fitness checks, framework init logs, and any warmup output. A snapshot restore is silent until the first Jobs in queue: 1 line, then jumps straight to your handler’s request-processing logs. The presence or absence of the boot sequence is the cleanest signal.

Is FlashBoot the same as RunPod’s “Active Workers” tier?

No. Active Workers are a billing tier where you pre-commit to a number of workers that are always on, billed at a discount in exchange for the 24/7 commitment. FlashBoot is a free runtime optimization that applies to flex (scale-to-zero) workers. The two can be combined: an Active Worker on the same endpoint can also benefit from FlashBoot when it cycles, though for a worker that never goes idle there’s nothing to snapshot.

Will FlashBoot survive a Docker image rebuild?

No. Each image gets its own SHA, and FlashBoot snapshots are scoped to (host, image SHA). When you push a new image, all existing snapshots are invalid. The first request after a rebuild on any host pays the full cold-start cost (image pull + warmup). Once each host has served the new image once, subsequent restores work normally.

What’s next

The runpod-mineru repo wraps all of this into one Docker image: MinerU 3.2.x + the MinerU2.5-Pro-2605-1.2B VLM, the JobScaler-bypass composition for warmup, structured logging, and the rest. Open source (GitHub), MIT-licensed, deploys from the RunPod Hub in two clicks.

If you want the deeper breakdown of which phases of a cold start cost what, the troubleshooting guide has the per-phase timing table from the same test runs. The scaling guide covers when to pair FlashBoot with workers_min ≥ 1 for fully predictable latency.

Disclosure: RunPod links in this post use a referral code that credits me at no cost to you. The post would read the same without it.