[Cherry-pick] PRs #2359 #2202 #2317 #2289 #2314 #2371 #2403 #2405 #2319 #2203 #2413 #2412 #2231 #2439 #2436 #2432 #2410 #2414 #2460 #2451 #2336 #2471 #2469 #2480 #2452 #2438 #2402 #2434 - #2502

Closed

chadvoegele wants to merge 28 commits into

release/0.47.0from

cherry-picks/release-0.47.0

Commits on Sep 22, 2026

Speed up FSDP2 MoE calibration by dropping redundant expert-weight gathers (#2359)

authored and
chadvoegele
committed
613175a
feat(quantization): PTQ support for Step-3.7 MoE checkpoints (#2202)

authored and
chadvoegele
committed
5d4d395
[6410139] Fix ONNX AutoCast for large external initializers (#2317)

ajrasane
authored and
chadvoegele
committed
839fba6
specdec: config_overrides for nested text_config checkpoints + load VLM-capable bases in merge_lora (#2289)

authored and
chadvoegele
committed
7800218
[6508436] Fix BF16 FP8 ONNX export (#2314)

authored and
chadvoegele
committed
d19c7d6
deprecate trtllm-build in weight_sparsity (#2371)

noeyy-mino
authored and
chadvoegele
committed
2fbcca2
Fix protobuf size-check failures in ONNX deployment (#2403)

noeyy-mino
authored and
chadvoegele
committed
68db2c9
Document the MSE calibration API (#2405)

authored and
chadvoegele
committed
aca3f13
Fix TEGroupedMLP quantizer checkpoint resharding (#2319)

jenchen13
authored and
chadvoegele
committed
c51568c
feat(quantization): fail fast when a quant config matches no weight quantizer (#2203)

authored and
chadvoegele
committed
e75d6c1
[6701308][OMNIML-5805] Correct ONNX PTQ documentation contracts (#2413)

authored and
chadvoegele
committed
83876db
[6463897] Fix narrow FP16 histogram calibration (#2412)

authored and
chadvoegele
committed
5ecede0
Fix distributed AutoQuantize scoring and share backward setup (#2231)

authored and
chadvoegele
committed
72a53b0
Document the nvfp4_act_headroom calibration variant in ptq.md (#2439)

authored and
chadvoegele
committed
60638ec
docs: replace legacy TensorRT-LLM engine deployment guidance (#2436)
chadvoegele
committed
9880ff9
Fail fast on non-finite AutoQuantize output gradients (#2432)

meenchen
authored and
chadvoegele
committed
6ea33c5
fix(specdec): resolve the eagle aux-layer preset in the vLLM hidden-state dump (#2410)

authored and
chadvoegele
committed
763cf95
Fix vLLM fakequant calibration for hybrid attention models (#2414)

kinjalpatel27
authored and
chadvoegele
committed
3d9872f
docs(recipes): point chat_template at where the templates actually live (#2460)

authored and
chadvoegele
committed
961741c
[6771663] Preserve ONNX API output types when wiring casts (#2451)

authored and
chadvoegele
committed
0470b12
[5565357] Fix SDXL NVFP4 export and performance (#2336)

authored and
chadvoegele
committed
a2cee6a
Move Nemotron Nano offline KD example (#2471)

jenchen13
authored and
chadvoegele
committed
a48fd62
docs: clarify canonical pruning documentation source (#1871) (#2469)

DIYA73
authored and
chadvoegele
committed
415e826
Fix hf_ptq.py discarding completed PTQ run on sanity-generate() failure (#2480)

authored and
chadvoegele
committed
39a61af
Fix hybrid stack spec serialization in Megatron-Bridge checkpoints (#2452)

authored and
chadvoegele
committed
b935126
Fix HF export crash when a dynamic-block quantizer has zero amax (#2438)

authored and
chadvoegele
committed
042e1ff
Noeyy/fix bug 6701777 (#2402)

noeyy-mino
authored and
chadvoegele
committed
6412746
[https://nvbugspro.nvidia.com/bug/6778095] Fix fused P-QDQ to respect disabled quantization during calibration (#2434)

yingguo-trt
authored and
chadvoegele
committed
eb22ef0

Navigation Menu

[Cherry-pick] PRs #2359 #2202 #2317 #2289 #2314 #2371 #2403 #2405 #2319 #2203 #2413 #2412 #2231 #2439 #2436 #2432 #2410 #2414 #2460 #2451 #2336 #2471 #2469 #2480 #2452 #2438 #2402 #2434 - #2502

[Cherry-pick] PRs #2359 #2202 #2317 #2289 #2314 #2371 #2403 #2405 #2319 #2203 #2413 #2412 #2231 #2439 #2436 #2432 #2410 #2414 #2460 #2451 #2336 #2471 #2469 #2480 #2452 #2438 #2402 #2434#2502
chadvoegele wants to merge 28 commits into
release/0.47.0from
cherry-picks/release-0.47.0

Commits on Sep 22, 2026

Speed up FSDP2 MoE calibration by dropping redundant expert-weight gathers (#2359)

feat(quantization): PTQ support for Step-3.7 MoE checkpoints (#2202)

[6410139] Fix ONNX AutoCast for large external initializers (#2317)

specdec: config_overrides for nested text_config checkpoints + load VLM-capable bases in merge_lora (#2289)

[6508436] Fix BF16 FP8 ONNX export (#2314)

deprecate trtllm-build in weight_sparsity (#2371)

Fix protobuf size-check failures in ONNX deployment (#2403)

Document the MSE calibration API (#2405)

Fix TEGroupedMLP quantizer checkpoint resharding (#2319)

feat(quantization): fail fast when a quant config matches no weight quantizer (#2203)

[6701308][OMNIML-5805] Correct ONNX PTQ documentation contracts (#2413)

[6463897] Fix narrow FP16 histogram calibration (#2412)

Fix distributed AutoQuantize scoring and share backward setup (#2231)

Document the nvfp4_act_headroom calibration variant in ptq.md (#2439)

docs: replace legacy TensorRT-LLM engine deployment guidance (#2436)

Fail fast on non-finite AutoQuantize output gradients (#2432)

fix(specdec): resolve the eagle aux-layer preset in the vLLM hidden-state dump (#2410)

Fix vLLM fakequant calibration for hybrid attention models (#2414)

docs(recipes): point chat_template at where the templates actually live (#2460)

[6771663] Preserve ONNX API output types when wiring casts (#2451)

[5565357] Fix SDXL NVFP4 export and performance (#2336)

Move Nemotron Nano offline KD example (#2471)

docs: clarify canonical pruning documentation source (#1871) (#2469)

Fix hf_ptq.py discarding completed PTQ run on sanity-generate() failure (#2480)

Fix hybrid stack spec serialization in Megatron-Bridge checkpoints (#2452)

Fix HF export crash when a dynamic-block quantizer has zero amax (#2438)

Noeyy/fix bug 6701777 (#2402)

[https://nvbugspro.nvidia.com/bug/6778095] Fix fused P-QDQ to respect disabled quantization during calibration (#2434)