fix(trt_llm): BEI max_num_tokens upgrade is a no-op

[chuenchen309]

What

_bei_specfic_migration logs that it is raising max_num_tokens to the BEI minimum, but the value never changes:

TrussTRTLLMBuildConfiguration(base_model="encoder", max_num_tokens=1024, checkpoint_repository=...)
# WARNING  truss.base.trt_llm_config: build.max_num_tokens=1024, upgrading to 16384
# .max_num_tokens -> 1024

The upgrade rebinds the local name instead of assigning the field:

self = self.model_copy(update={"max_num_tokens": BEI_REQUIRED_MAX_NUM_TOKENS})

model_copy returns a new object; rebinding self inside model_post_init discards it when the method returns.

Why this looks like a slip rather than intent

• The next two lines of the same method assign fields directly and do stick:

  self.plugin_configuration.paged_kv_cache = False
  self.plugin_configuration.use_paged_context_fmha = False

  One method, two behaviours.
• The only other model_copy in the file assigns the result back to a field:

  self.runtime = self.runtime.model_copy(update={"webserver_default_route": route})

• The logger.warning announces an upgrade that then doesn't happen.

Impact

serving_image_builder.py notes it is "enforcing max_num_tokens here and in engine-builder". The engine-builder half is what silently doesn't apply: the router is configured for the BEI minimum while the engine is built for the user's lower value.

Fix

Assign the field. Verified across the paths:

base_model	input	after
encoder	1024	16384 ✅
encoder	8192 (default)	16384 ✅
encoder	16384	16384 ✅
decoder	1024	1024 (unchanged) ✅

The adjacent paged_kv_cache = False still applies.

Tests

Two cases: the encoder path upgrades, the decoder path does not. The encoder one fails on main and passes here. grep max_num_tokens truss/tests/ previously returned nothing — the field had no coverage, which is why this went unnoticed.

• pytest truss/tests/trt_llm/ — 23 passed, 1 skipped
• Full non-Docker suite — 1090 passed (1088 baseline + 2 new), no regressions
• uv run ruff check / ruff format --check clean, per AGENTS.md

---

AI disclosure: drafted with Claude Code (Opus 4.8), including the root-cause trace and the tests. I ran the repro and the suite locally and reviewed the diff before opening.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[cretz]

Thanks for the contribution! One question below and also CLA will need to be signed (click link above), thanks!

[cretz]

Thanks for the fix!

Can you help me understand how you came across this bug? Specifically the steps you took on the platform to trigger it?