[Cherry-pick] PRs #1393 #1389 #1268 #1397 #1402 #1411 #1410 #1419 #1408 #1416#1426

Merged

kevalmorabia97 merged 11 commits into

release/0.44.0from

cherry-picks/release-0.44.0

May 11, 2026

Commits on May 11, 2026

llm_sparsity: Set warmup_steps 0 instead of 0.0 for transformers 5.x compat (#1393)
kevalmorabia97
committed
[specdec_bench] Stratify --num_requests across categories (#1389)

milesial
authored and
kevalmorabia97
committed
Centralize 'trtexec' subprocess runs in ONNX into a single function (#1268)

gcunhase
authored and
kevalmorabia97
committed
[Quantization] Saturate NVFP4 export FP8 scale cast to avoid NaN (#1397)

cjluo-nv
authored and
kevalmorabia97
committed
[NVBug 6102977] Disable use_cache during export dummy forward for custom configs (#1402)

meenchen
authored and
kevalmorabia97
committed
[NVbug 6142360] Share fused gate_up amax fallback to keep weight_scale_2 consistent (#1411)

meenchen
authored and
kevalmorabia97
committed
[NVBug 6143871] Fix awq_lite uncalibrated branch leaving input_quantizer disabled (#1410)

cjluo-nv
authored and
kevalmorabia97
committed
Fix 6115282 (#1419)

sugunav14
authored and
kevalmorabia97
committed
[6106576] Restore qdq_utils symbols used by edgellm 0.6.1 (#1408)

ajrasane
authored and
kevalmorabia97
committed
fix(llm_eval): migrate lm_eval_hf.py to lm-eval >= 0.4.10 HarnessCLI (#1416)
kevalmorabia97
committed
Pin deepspeed<0.19 for older torch compatibility
kevalmorabia97
committed