Integrate Miles fused NVFP4 fake-QAT QDQ

[zianglih]

• I, the PR author, have personally reviewed every line of this PR.

What does this PR do?

@HumansAnd

Adds the minimal Megatron side of a paired Miles/Megatron split for fused NVFP4 fake-QAT QDQ.

• Scope: four lines in TEGroupedLinear._get_weight_tensors() gate on
  OPEN_TRAINING_NVFP4_FAKE_QAT_FLAG=1, lazily import the Miles-owned adapter, and transform
  TE's existing discrete weight list before the current INT4-QAT hook.
• Disabled path: Megatron does not import Miles or CuTe DSL when the flag is off.
• Ownership: the CuTe kernel, FP32 amax/STE wrapper, full numerical matrix, real grouped-linear
  forward/backward test, and benchmark live in the paired Miles PR. This PR adds no duplicate
  kernel or unit-test files.
• Recipe premise: the enabled recipe targets TE 2.17+ discrete rank-2 weights. It deliberately
  carries no pre-2.17 fallback or runtime support validator; recipe compatibility is configured by
  the caller.
• History: preserves draft Fuse NVFP4 fake-QAT quantize-dequantize with CuTe DSL #88 unchanged as the complete single-repository implementation and
  re-expresses it as a cleaner ownership split. This remains a follow-up to [WIP] Add NVFP4 fake QAT for grouped experts #75.

Paired Miles kernel/test PR:
radixark/miles#2864

Issue tracking

Linked issue: N/A (paired fork refactor of draft #88; related to #75)

Validation

• Megatron commit: 65940a7197427997bbfd411b3944849f68c1aaf0
• Megatron base: c160f0717e40f052de740026cf62739340e7b970 (miles-main)
• Paired Miles head: ca0c3425f2c5b5f8ec97fc4af6e8fd976d32f704
• C2 validation/benchmark commit: ffa7dc4ac3741c1e4f2c896aa78273730c418fe4
  (the head differs only by restoring the pre-existing CI registration metadata)
• Image: radixark/miles:dev-202609010119
  (sha256:3ee2af22223ee8f0d0b0bf19340fc06bf8da6ddb26a36c76136f8f5e5c462728 on amd64/B300)
• Hardware/runtime: C2, 8x NVIDIA B300 SXM6 AC (SM103); PyTorch 2.13.0+cu130, CUDA 13.0,
  Transformer Engine 2.17.0, Cutlass DSL 4.6.2

The Miles Blackwell file includes the real TEGroupedLinear hook, three discrete weights,
bit-exact TE QDQ comparison, forward, and identity-STE backward. The full run uses two visible GPUs
so the non-current-device restoration case executes rather than skips.

cd /hai-workspace/nvfp4-qdq-split/miles
CUDA_VISIBLE_DEVICES=0,1 PYTHONWARNINGS=ignore \
PYTHONPATH=/hai-workspace/nvfp4-qdq-split/miles:/hai-workspace/nvfp4-qdq-split/Megatron-LM \
pytest -q -o addopts= tests/fast-gpu/test_nvfp4_quantizer.py

........................................................................ [  4%]
........................................................................ [  9%]
........................................................................ [ 13%]
........................................................................ [ 18%]
........................................................................ [ 22%]
........................................................................ [ 27%]
........................................................................ [ 31%]
........................................................................ [ 36%]
........................................................................ [ 41%]
........................................................................ [ 45%]
........................................................................ [ 50%]
........................................................................ [ 54%]
........................................................................ [ 59%]
........................................................................ [ 63%]
........................................................................ [ 68%]
........................................................................ [ 73%]
........................................................................ [ 77%]
........................................................................ [ 82%]
........................................................................ [ 86%]
........................................................................ [ 91%]
........................................................................ [ 95%]
................................................................         [100%]
1576 passed in 20.68s

Local Megatron checks:

isort: passed
Black: 1 file left unchanged
pylint: 10.00/10
Ruff: All checks passed
compileall: passed
git diff --check: passed

tools/autoformat.sh was invoked but its fetch step cannot resolve the fork-only miles-main
branch on its hardcoded NVIDIA remote. The equivalent isort, Black, pylint, and Ruff commands were
run individually against the changed file and produced the results above; this was an
infrastructure/setup failure rather than a formatter failure.

The paired eight-weight [4096,6144] benchmark reports 1.623x overall geomean speedup versus
native TE quantize-dequantize (1.635x BF16, 1.611x FP16); complete raw timing is owned by the
Miles PR.

Contribution process

Pre-checks

• I have added relevant unit tests (in the paired Miles Blackwell test)
• I have added relevant functional tests (real paired grouped-linear forward/backward)
• I have added proper typing to my code Typing guidelines
• I have added relevant documentation
• I have run the autoformatter.sh checks on my PR

Code review

Feel free to message or comment @NVIDIA/mcore-oncall to help accelerate your merge into main. The less complex your PR is, the faster it will be approved and merged!

All PRs start as draft. If you open a non-draft PR, it will be automatically converted to draft.

Step 1: Mark PR as "Ready for Review"

1. When the PR is ready, click Ready for Review.
2. An oncall reviewer is auto-assigned and expert reviewers are notified based on the changes.

⚠️ Only mark as ready once merge conflicts are resolved and CI is passing.

Step 2: Final Review

For changes under megatron/core, expert approvals lead to the Final Review label and final review.

Step 3: Approved

The Approved label is applied automatically after the required reviews.

Merge

Any member of mcore-engineers can merge the PR.