[Bugfix][Qwen4Exp] Detect the FP8 PLE table from ModelOpt mixed-precision configs

[Peuqui]

Purpose

nvidia/Qwen3.8-Flash-Next-NVFP4 stores its PLE table exactly like the other
Flash-Next NVFP4 exports: 129 FP8 E4M3 shards under
model.language_model.layers.1.ple.ple_embedding.ngram_embedding with one
weight_scale. It declares that table in the ModelOpt mixed-precision config
(quantized_layers[...ngram_embedding] = {"quant_algo": "FP8"}) and does not
set ple_embedding_dtype in text_config.

_get_ple_embedding_quant_method only recognises FP8 storage through
ple_embedding_dtype == "float8_e4m3fn" or an Fp8Config. For this checkpoint
it returns None, and on pre-Ampere cards, where the pinned-host PLE placement
is taken, the engine refuses to start:

NotImplementedError: Qwen4Exp pinned-host PLE requires FP8 checkpoint storage

The change adds a ModelOptMixedPrecisionConfig branch: if the table is not
excluded and its per-layer algorithm resolves to FP8, the existing
Qwen4ExpPLEFp8EmbeddingMethod is selected. The lookup goes through the
config's own is_layer_excluded and _resolve_quant_algo, which already map
between the checkpoint's model.language_model. and vLLM's language_model.model.
prefixes and are used the same way for MTP experts in mtp_fp8_experts.py.
Checkpoints that set ple_embedding_dtype or use Fp8Config take the same path
as before.

Not a duplicate: searched open and closed PRs and issues for "PLE FP8",
"ple_embedding_dtype", "ModelOptMixedPrecision PLE", "pinned-host PLE",
"Flash-Next nvidia". #553 concerns the MTP experts of the same checkpoint and
does not touch the PLE table.

Test Plan

.venv/bin/python -m pytest tests/models/qwen4_exp/test_ple.py
pre-commit run --files vllm/models/qwen4_exp/nvidia/ple_layer.py tests/models/qwen4_exp/test_ple.py
pre-commit run mypy-3.10 --hook-stage manual --files vllm/models/qwen4_exp/nvidia/ple_layer.py tests/models/qwen4_exp/test_ple.py

End to end with nvidia/Qwen3.8-Flash-Next-NVFP4 revision fc694b54 on 2x Quadro
RTX 8000 + 2x Tesla V100, TP2 x PP2, pinned-host PLE with a 6 GiB host budget.

Test Result

• Tests: 59 passed, 3 skipped. The new test builds the config from a ModelOpt
  mixed-precision dict and checks both prefix spellings: FP8 table selects the
  FP8 embedding method, an unlisted table and an excluded table do not. Without
  the fix both parametrized cases fail.
• pre-commit and mypy-3.10 pass.
• Before: startup fails on the PLE stage with the NotImplementedError above.
• After: the checkpoint loads straight from its Hugging Face snapshot, without
  editing its config, and serves; the same run is described in [Bugfix][Qwen4Exp] Allow checkpoint FP8 MTP experts under pipeline parallelism #639
  for FP8 MTP experts under pipeline parallelism, which this checkpoint needs as
  well for MTP on that topology.

These end-to-end runs were made on our fork (1Cat main 80c88e8 plus our open
PRs).

AI assistance (Claude) was used for this change and this description. I reviewed
every changed line and ran the tests and measurements above.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

🤖 Generated with Claude Code