Keep the newer-mapper probe alive when the fetched mapper has no fp8 tables

[danielhanchen]

Follow-up to #7478.

_get_new_mapper reads the two fp8 tables out of the fetched mapper.py under that file's own names, unlike the three NEW_ names it renames itself. A mapper.py that does not define them raises KeyError, the bare except swallows it, and the function returns five empty dicts, so the 4bit and 16bit upgrade check stops firing too. That check is the reason the probe exists, and it is the half that has nothing to do with fp8.

The probe fetches from main at runtime, so the file it gets is not under the installed version's control. Every mapper.py older than the fp8 tables is one that does not define them. Serving each version that has actually existed in this repo, and asking whether the 4bit upgrade check still works:

fetched mapper.py	date	before #7478	after #7478	this PR
d6bb89a (predates the fp8 tables)	2025-11-07	alive	dead [0, 0, 0, 0, 0]	alive [400, 997, 591, 0, 0]
86f7080	2025-11-25	alive	alive	alive
f840119	2026-03-03	alive	alive	alive
800ddc9	2026-04-15	alive	alive	alive
current main	2026-07-27	alive	alive	alive

Before #7478 this degraded gracefully, because a missing name was only ever a NameError on a name the code did not need. Reading the two names with .get restores that: the 4bit half keeps working and only the fp8 half goes empty, which costs nothing real, since the probe runs only after the installed tables have already missed.

Test

tests/test_new_mapper_fetched_fp8.py, two cases:

1. a fetched mapper.py with no fp8 tables must not take the 4bit half down. Fails on current main with assert ({}), passes here.
2. an fp8 repo that only the fetched mapper knows must still raise the upgrade NotImplementedError, and answering it must not adopt the fetched tables.

The second case covers a gap in test_new_mapper_no_global_leak.py: that file serves the repo's own mapper.py as both the installed and the fetched source, so fp8_block is not block is satisfied by any fresh dict, since exec always allocates new ones. It pins allocation, not provenance. A _get_new_mapper that returns the fetched fp8 tables filtered down to keys the installed tables already have, which is functionally a revert of 881180f, passes that file 2/2. This test splices an extra "8" entry into the fetched source only, and stubs importlib so vllm looks absent, since from 0.12.0 on __get_model_name returns the original name on the first pass and the fp8 path is never reached.

Against the three relevant revisions:

revision	result
before #7478	fails, installed table was rebound
797f5e2 (isolate only)	fails, DID NOT RAISE
current main	fails case 1
this PR	passes

tests/test_new_mapper_fetched_fp8.py::test_probe_answers_for_an_fp8_repo_only_the_fetched_mapper_knows PASSED
tests/test_new_mapper_fetched_fp8.py::test_probe_survives_a_fetched_mapper_without_the_fp8_tables PASSED
tests/test_new_mapper_no_global_leak.py .. PASSED
tests/test_bad_mappings_redirect.py . PASSED
tests/test_gemma_2b_mapper_key.py . PASSED
tests/test_get_model_name.py .. PASSED (40 subtests)

Full CPU suite: 2940 passed, 56 skipped. The test is CPU-only and makes no network calls: it ast-extracts the resolvers and stubs requests with the repo's own mapper.py, matching tests/test_bad_mappings_redirect.py.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.