Read the dtype from text_config in convert_ggml_to_int8.py - #2946

…and QAD (#2915)

### What does this PR do?

[2/2], after #2917 (date-prefixed announcement filenames, merged).

Type of change: documentation

**Executive summary**

- **IQ2_S + NVFP4 matches FP8 with no training.** It uses IQ2_S on the
experts' gate and up projections and NVFP4 on their down projections,
3.21 bits per expert weight. The model is ~155 GB, against 328 GB for
FP8 (2.1x smaller), and its six-benchmark mean is within 0.1 points of
FP8.
- **IQ2_S is the smallest option.** At 2.56 bits per weight on all
expert weights, the model is ~130 GB (2.5x smaller), and its mean is 2.3
points below FP8. Most of the loss is runaway generation on long tasks.
- **QAD fixes long-context runaways.** 300 iterations on IQ2_S brought
AA-LCR from 70.0 to 74.0 (FP8 74.9) and cut answers hitting the token
budget from 6.9% to 0.6%. The mean gap shrank from 2.3 to 1.4 points.
- **We recommend the PTQ recipes.** QAD cost ~4,500 GPU-hours, didn't
restore knowledge recall, and IQ2_S + NVFP4 does better without
training. Both recipes still trail FP8 by ~2–3 points on AA-Omniscience.
Evaluations served BF16 fake-quant exports; IQ2_S can also be converted
to an INT8 weight-only checkpoint that vLLM serves at 86 instead of 148
GiB of weights per GPU (#2903).

Adds an announcement page (blog) for the GLM-5.3-Flash IQ2 study, with a
results figure in the Qwen3.6 announcement's style, a card on the
announcements index and a Latest News entry in the README. It presents
two PTQ recipes as alternatives:

| | FP8 checkpoint | IQ2_S | IQ2_S + NVFP4 |
|---|---|---|---|
| Routed experts | FP8 | IQ2_S | IQ2_S gate/up, NVFP4 down |
| Routed-expert bits per weight | 8 | 2.56 | 3.21 |
| Model size | 328 GB | ~130 GB | ~155 GB |
| Mean delta vs FP8, six benchmarks | – | −2.3 | −0.1 |

It also covers knowledge recall (HLE, AA-Omniscience), RULER-1M, QAD on
the IQ2_S model (300 iterations, −1.4, with worse recall), the cost, and
the GLM-specific reproduction steps.

**Blog, not a full tutorial.** QAD did not become the recipe: IQ2_S +
NVFP4 reaches FP8 with PTQ alone. So the page points to the
Qwen3.6-35B-A3B tutorial for the general flow and lists only GLM's
differences: the recipes, `--no_moe_grouped_gemm`, EP16 PTQ and
fake-quant export, long-reasoning calibration, eval settings, and the
QAD flags.

**Baseline:** every delta is against the FP8 release
(`zai-org/GLM-5.3-Flash`), not the BF16 checkpoint. A note near the top
says so.

**Sizes:** computed from each tensor's bits per weight using the FP8
checkpoint's safetensors headers, since evaluation used BF16 fake-quant
exports. The post also covers serving IQ2_S as INT8 weight-only
(`convert_ggml_to_int8.py`, #2903 / #2946): accuracy, weight memory and
vLLM throughput with the Humming kernels. IQ2_S + NVFP4 has no such path
yet.

**Dependencies:** none for merging this docs-only change. Reproducing
the experiments needed:
- ModelOpt: #2873 (merged; GGML PTQ, QAD and fake-quant export for large
MoE/VLM models) and #2539 (GLM-5.3-Flash support for Megatron-Bridge PTQ
and HF export, plus both IQ2_S recipes).
- Megatron-Bridge: NVIDIA-NeMo/Megatron-Bridge#6044 (GLM-5.3-Flash
bridge).
- Megatron-LM: #7521 (mHC precision controls) and #7522 (KPool DSA
indexer), both merged, and the open #7054 (KDA hybrid layers), plus two
training fixes for QAD: ceaa6c5d (#5835, distributed-optimizer save/load
for mixed fp32/bf16 param groups) and bfb17738 (#7522, freeze the DSA
indexer when `dsa_indexer_loss_coeff` is 0).
- transformers >= 5.16.1 (`glm5_next`), nemo:26.10 base image, vLLM >=
0.30 for evaluation.

The recipes stay in #2539: they use Megatron-Bridge module names, and
`examples/hf_ptq` would neither apply GPTQ to the fused HF routed
experts nor match the dense-MLP and NVFP4-down patterns, so they cannot
be used without the Megatron-Bridge support.

**Prefix-caching check (done):** our evals serve vLLM 0.31 with prefix
caching on, which on vLLM < 0.31 corrupts GLM's long-context cache
(vllm#57477). An A/B on FP8 AA-LCR (8 runs each) gives 74.88 ± 0.87 with
caching on and 76.12 ± 0.83 with it off: about one standard error apart,
so the 0.31 fix holds and the published numbers stand.

### Usage

N/A (documentation)

### Testing

- Parsed the new page with docutils: no errors or warnings.
- Did not run a full Sphinx build.

### Before your PR is "*Ready for review*"

- Is this change backward compatible?: N/A
- If you copied code from any other sources or added a new PIP
dependency, did you follow guidance in `CONTRIBUTING.md`: N/A
- Did you write any new necessary tests?: N/A
- Did you update
[Changelog](https://github.com/NVIDIA/Model-Optimizer/blob/main/CHANGELOG.rst)?:
N/A
- Did you get Claude approval on this PR?: ❌

### Additional Information

Dated 2026-10-28, the publish date. The Reproducing section points at
Megatron-LM main (which now has #7521) with #7054 merged in, and at
Megatron-Bridge #6044; it will be updated once both merge.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Signed-off-by: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>