Add Nemotron 3.5 Lightning DFlash serving (stacked on #248)

[davidtai]

Stack dependency

This PR is stacked on #248, which supplies the Glimmer DFlash backend used as the integration base. Until #248 merges, GitHub will show both commits in this PR. The Nemotron-specific change is dc68a6c0.

What changed

Adds native MTPLX serving support for NVIDIA Nemotron 3.5 Lightning 30B-A3B with NVIDIA's released DFlash drafter. The change adds a streaming NVIDIA ModelOpt NVFP4/mixed-FP8 to MLX affine-4 conversion path, maps the released DFlash config and weights onto the existing Glimmer adapter, installs a fixed construction-time capture route for Nemotron-H's hybrid Mamba/attention trunk, and integrates DFlash pair inspection, backend metadata, server defaults, telemetry, and tests.

Why

NVIDIA publishes the Lightning target and DFlash assistant in ModelOpt formats and with a hybrid Nemotron-H target topology. The existing Glimmer adapter establishes the block-diffusion algorithm, but the Lightning checkpoint needs its own quantization conversion, direct layer-output taps, YaRN/config mapping, and hybrid-cache rollback/repair to preserve exact greedy target-prefix verification.

How to verify

uv run --frozen --with pytest pytest -q \
  tests/test_modelopt_nvfp4.py \
  tests/test_dflash_nemotron.py \
  tests/test_artifacts.py::test_dflash_pair_bundle_inspects_as_native_runtime \
  tests/test_artifacts.py::test_gemma4_pair_bundle_inspects_as_assistant_runtime \
  tests/test_compressed_tensors.py \
  tests/test_server_openai.py::test_gemma4_request_uses_draft_block_before_depth

uv run mtplx inspect <dflash-pair-bundle> --require-mtp --json
uv run mtplx serve --model <dflash-pair-bundle> --mtp

Local real-model verification used:

• target nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4@e0b753dc24903ad4d62f5696077da22020eca89a
• drafter nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DFlash@bbea7f89d6221e2242a2f00f958afce59d32fab7
• 18 focused tests passed
• DFlash matched independent greedy target AR for 32/32 generated tokens
• standard mtplx serve returned HTTP 200 from health, models, and chat endpoints; the chat request recorded 112 drafted tokens, 16 accepted drafts, 16 verify calls, and requested block size 8

Notable decisions

• Preserve NVIDIA's released block size 8 as a construction-time invariant; generic native-MTP depth defaults do not resize the installed DFlash lane per request.
• Validate topology, capture layers, and quantization layout during construction and conversion, not inside the measured per-layer path.
• Requantize ModelOpt NVFP4/mixed-FP8 weights to standard MLX affine 4-bit while retaining pinned source revisions in the converted configs.
• Use snapshot/rollback plus committed-prefix repair for Nemotron-H's hybrid recurrent/attention cache after partial acceptance.
• Keep the initial serving contract greedy-only. No non-greedy or performance-speedup claim is made.
• Skip retokenized SessionBank postcommit for DFlash until this external backend exposes a safe final-state snapshot contract.

[youssofal]

Stack noted. Review lands after #248 does; the dc68a6c split makes the Nemotron leg easy to isolate.