[HANDS-ON BUG] First_agent_template copies fail with 'No @spaces.GPU function detected' when set to ZeroGPU — 39-59% of July/August 2026 copies

[ashishsinha1602]

Describe the bug

Copies of First_agent_template made since July 2026 are failing at startup at a sharply elevated rate, with an error unrelated to the smolagents problem in #520. The cause appears to be students selecting ZeroGPU hardware for a template that has no @spaces.GPU function.

Evidence

I have a census of all 1,458,692 public Spaces (August 2026) and can isolate every copy of this template. There are 28,689 of them, one per account.

Broken rate by the month the copy was made:

copy made	n	in an error state
2026-04	1,184	3.7%
2026-05	1,077	4.0%
2026-06	1,104	4.1%
2026-07	477	39.0%
2026-08	281	58.7%

That is a roughly tenfold jump against the preceding three months.

It is not transient. I re-queried 40 of the affected Spaces live in September, a month after the census: 40 of 40 were still in RUNTIME_ERROR.

The error is specific. Of 78 recently-broken copies re-queried, 76 returned an error message, and 67 of those (88.2%) were the same one:

No @spaces.GPU function detected during startup

Only 4 of 76 mentioned smolagents or CodeAgent, so this is not #520 resurfacing.

The mechanism. Comparing recent copies by hardware:

recent copies	hardware
broken	33/35 on zero-a10g
not broken	21/35 on cpu-basic, 5 on zero-a10g

The upstream template README specifies no hardware, so a copy defaults to free CPU and works. app.py is CPU-only smolagents code with no @spaces.GPU decorated function. When a learner switches their Space to ZeroGPU, which is an understandable thing to do when building an "AI agent", startup fails with the error above.

Suggested fix

A line in the hands-on instructions saying the template runs on the free CPU tier and that ZeroGPU must not be selected would prevent it. Alternatively, the error is common enough now to be worth a troubleshooting entry.

Scale

351 copies created in July and August 2026 are currently broken. Across all cohorts, 3,975 accounts hold a broken copy of this template.

Method and reproducibility

Detection and analysis code: https://github.com/ashishsinha1602/dataset-integrity-audit

Copy-relationship data: https://huggingface.co/datasets/Ashsinha1/hf-spaces-clones

Caveats: stage is a snapshot and drifts; SLEEPING is not counted as broken; the live re-queries above are samples, with sizes stated.

[ashishsinha1602]

Correction to my own report above: the cause is not learners choosing ZeroGPU. They appear to have no free choice left.

I attributed the failures to students selecting ZeroGPU voluntarily. Reading the Space's own discussion threads, that is wrong, and the real mechanism is worse.

Multiple independent reports on agents-course/First_agent_template say CPU Basic is no longer selectable when duplicating:

• discussion #733 (2026-08-14, 7 comments): "Why I can't duplicate this space for free as shown in the lecture example?", followed by "the CPU basic version is not available" and "the duplicate button is grayed out and no free cpu version" from five further accounts.
• discussion #736 (2026-08-25): "on the tutorial page it shows that I can duplicate the space for free. Now I see that CPU Basic requires a Pro subscription."
• discussion #737 (2026-08-26): the resulting No @spaces.GPU function detected during startup failure, with two more learners reporting the same and one saying they cannot continue the course.

That fits the measurements in my original report far better than voluntary selection does. If a learner must pick ZeroGPU because free CPU is unavailable, and the template has no @spaces.GPU function, then every copy made after that change fails — which is what the cohort data shows:

copy made	n	in an error state
2026-04	1,184	3.7%
2026-05	1,077	4.0%
2026-06	1,104	4.1%
2026-07	477	39.0%
2026-08	281	58.7%

and it explains why 33 of 35 sampled broken recent copies are on zero-a10g while working ones are mostly cpu-basic: the working ones were duplicated before the change.

This reframes the fix. A documentation note saying "do not select ZeroGPU" is useless if ZeroGPU is the only option a free account is offered. What would actually resolve it is one of:

1. restoring free CPU Basic as a duplication target for this template, or
2. adding a trivial @spaces.GPU decorated function to the template so a ZeroGPU copy starts, or
3. updating Unit 1 to reflect that duplication is no longer free, since the tutorial still shows it as such.

One learner has already worked around it by abandoning Spaces entirely and running locally against Ollama and a free Gemini key (#733), which is a reasonable answer but not the course as written.

I cannot verify the hardware-availability change directly from the API, so that part rests on the learner reports rather than my own measurement. The cohort breakage numbers and hardware split are mine and are reproducible.

[ashishsinha1602]

This affects a second unit, so the blast radius is larger than my report above.

I applied the same measurement to Final_Assignment_Template (Unit 4). It shows the same pattern, and the hardware separation is cleaner than for Unit 1.

3,113 copies exist. Of the 472 created since 2026-06, 88 are in an error state. Sampling them live:

recent copies sampled	hardware
broken (35)	26 zero-a10g, 9 cpu-basic
not broken (35)	35 cpu-basic, 0 ZeroGPU

Every working recent copy in the sample is on free CPU. Of the 35 broken ones, 22 return No @spaces.GPU function detected during startup — the same startup check as Unit 1. Only 3 mentioned smolagents, so this is not #520.

For the hardware shift itself, sampling copies of the Unit 1 template by month created and re-querying hardware live:

copies created	on cpu-basic
2026-06	30 of 30 (100%)
2026-07	11 of 30 (36.7%)
2026-08	2 of 30 (6.7%)

Free CPU has not been withdrawn outright — a small share of copies still get it — but it went from universal to rare across two months, and the breakage of both templates tracks that change.

Combined current impact across the two templates: First_agent_template has 351 copies from 2026-07/08 in an error state, Final_Assignment_Template has 88 since 2026-06. Learners are hitting it in the Space discussions rather than here — see #733 (six accounts reporting they cannot duplicate for free) and #737 (three reporting the resulting startup error, one saying they cannot continue the course).

Since both templates are CPU-only code with no GPU function, adding a trivial @spaces.GPU stub to each would make ZeroGPU copies start, and would not change behaviour for copies that still land on CPU. That is a smaller change than restoring the hardware option, if the hardware change was intentional.

As before: I can only observe what the API reports. I cannot tell whether the hardware shift is a pricing change, a quota change, a staged rollout, or a bug, and that distinction decides which fix is right.

[ashishsinha1602]

Confirmed cause. In my first report I said I could not verify the hardware change from the API and that the mechanism rested on learner reports. That is no longer true — the API states it directly.

Attempting to create a Space on a free account returns HTTP 402 with this body:

Static Spaces are free for everyone, but hosting Gradio and Docker Spaces on
free cpu-basic requires a PRO subscription. Subscribe at https://huggingface.co/pro

I ran it as a controlled test rather than relying on a single error, creating one Space per SDK on a free account:

SDK	result
gradio	refused, HTTP 402
docker	refused, HTTP 402
static	created successfully (probe deleted afterwards)

The static control succeeding is what makes this conclusive: it is not a rate limit, an outage, or an account problem.

This explains every measurement in the thread above. First_agent_template is a Gradio Space, so a learner on a free account can no longer put a copy on free CPU. ZeroGPU is what remains available to them, the template contains no @spaces.GPU function, and startup then fails with No @spaces.GPU function detected during startup. The cohort data matches exactly — copies on cpu-basic fell from 100% (June) to 36.7% (July) to 6.7% (August), while breakage rose from 4.1% to 39.0% to 58.7%.

Two things follow for the course.

The Unit 1 tutorial still presents duplicating the Space as free, which is what learners are reporting in discussion #736 and #733. That page needs updating regardless of what else is decided.

Adding a trivial @spaces.GPU stub to First_agent_template and Final_Assignment_Template would let ZeroGPU copies start, and would change nothing for copies that still land on CPU. That is a two-line fix and it does not depend on anyone reversing a pricing decision.

One inconsistency worth flagging separately: the public pricing page at https://huggingface.co/pricing still lists CPU Basic as FREE in its hardware table, with no indication that Gradio and Docker Spaces on it now require PRO. Both statements cannot be simultaneously useful to someone deciding whether they can follow this course for free.

I am not affiliated with Hugging Face. All checks above are reproducible with a free account.

[renukabodduru]

There is no option of free tier CPU ,if we go to Zero GPU that is costing us to PRO ,the CPU basic is also appearing to be PRO. How to fix this??

[ashishsinha1602]

@renukabodduru your observation matches the measurements in this thread, and it is not something wrong with your account.

Creating a Gradio Space on free CPU now returns HTTP 402 with this body, verified on a free account:

Static Spaces are free for everyone, but hosting Gradio and Docker Spaces on
free cpu-basic requires a PRO subscription.

I tested one Space per SDK: gradio refused (402), docker refused (402), static created successfully. So CPU Basic still exists as free hardware, but Gradio Spaces on it now need PRO — which is why the option looks unavailable to you, and why ZeroGPU is what remains.

What you can do, in order of least friction:

1. Run the agent locally. This is what @csongorosz did earlier in this thread, using Ollama or a free Gemini API key. It is the most reliable route today and it completes the Unit 1 exercise.
2. If your Space already exists on ZeroGPU, adding a stub makes the startup check pass, because the failure is only that no GPU function is present:

import spaces

@spaces.GPU
def _noop():
    return None

3. Check Settings once for CPU basic — a small number of copies still land on it (2 of 30 sampled from August), so it is worth a look before assuming.

I do not work for Hugging Face and cannot restore the option. The measurements are reported upstream in this issue so the course maintainers have them; the tutorial page still describes duplication as free, which is the part that most needs updating.

[ashishsinha1602]

@renukabodduru you are right that ZeroGPU is also gated, and my advice above was incomplete. Checking Hugging Face's own documentation rather than assuming, here is the full eligibility picture.

From the ZeroGPU docs, verbatim:

Free personal accounts: accounts in good standing (verified email, account older than 30 days) can host up to 2 ZeroGPU Spaces for free.

And from the create-Space API on a free account, verbatim:

Static Spaces are free for everyone, but hosting Gradio and Docker Spaces on free cpu-basic requires a PRO subscription.

Putting those together, for someone starting the course today:

route	available to a new free account?
Gradio on cpu-basic	No — requires PRO
ZeroGPU	No — requires the account to be over 30 days old
ZeroGPU, account 30+ days & verified email	Yes, up to 2 Spaces
Static Space	Yes, but cannot run this template

So a learner who signs up to take this course has no free way to host the Unit 1 template at all, and must wait 30 days, subscribe, or run locally. That is a stronger statement than anything earlier in this thread, and it explains why the options look uniformly unavailable to you rather than just CPU Basic.

I should correct my own earlier suggestion: adding an @spaces.GPU stub only helps someone who is able to host a ZeroGPU Space. If your account is under 30 days old, that fix does nothing for you, and I should have checked the eligibility rules before recommending it.

What actually works today for a new account: run the agent locally, as @csongorosz described earlier in this thread — Ollama, or a free Gemini API key. That path has no Hugging Face hosting requirement and completes the exercise.

For the maintainers, this sharpens the documentation issue considerably: the Unit 1 tutorial still presents duplicating the Space as free, and there is now no free hosting route for a newly created account. Whether or not the pricing change is reversed, the course text no longer matches what a new learner can actually do.

I do not work for Hugging Face; every claim above is quoted from public documentation or reproducible on a free account.