Cherry pick Habana software v1.21.0 commits#2180

Merged

XuehaoSun merged 48 commits intomasterfrom

cherry_pick_v1.21

May 13, 2025

Commits on Apr 22, 2025

[SW-205334][SW-187731] Reintroduce PR #67 with PatchedKVCache fix of [SW-214296] (#97)

authored and
xinhe3
committed
[SW-218081] temproray disable fp8_static_quant test (#131)

ulivne
authored and
xinhe3
committed
[FSW-12066] Add initial multi device support (#91)

ulivne
authored and
xinhe3
committed
[FSW-12066] small fixes in xpu quantized func (#145)

ulivne
authored and
xinhe3
committed
[SW-218197] fix bug in Mixtral unitscale (#139)

linoybu
authored and
xinhe3
committed
[ALGO-808] add support for int4 weights + fp8 activations - phase 1 (#43)

authored and
xinhe3
committed
[SW-218081] Re-enable tests (#140)

ulivne
authored and
xinhe3
committed
[SW-214378] remove creation of nc_workspace in each INC run (#151)

Yantom1
authored and
xinhe3
committed
[SW-219274] - Fix getting error log when lm_head in vLLM does not have measurements for scale calculation (#152)

nirda7
authored and
xinhe3
committed
[SW-207602] INC to Support fp8 communication in PatchedRowParallelLinear (#1)

authored and
xinhe3
committed
[SW-207602] Fix bug with PatchedRowParallelLinear (#158)

Tiefen-boop
authored and
xinhe3
committed
[SW-219745] fix fp8 GaudiMixtralSparseMoeBlock graph break (#161)

dudilester
authored and
xinhe3
committed
Blockwise gptq (#155)

Asaf Karnieli
authored and
xinhe3
committed
Raise error when measuring PC without shapes (#163)

HolyFalafel
authored and
xinhe3
committed
[SW-218303] Fix incorrect bias addition point in PatchedColumnParallelLinear (#160)

Tiefen-boop
authored and
xinhe3
committed
Correct `PatchedVLLMKVCache` to measure the whole input (#170)

authored and
xinhe3
committed
[SW-218484] Enhance log for saving (#134)

xin3he
and
xinhe3
committed
[SW-222320] Optimize code for TPC fuser in dynamic quantization (#173)

HolyFalafel
authored and
xinhe3
committed
[SW-221372] allow running Mixtral measurment phase using torch.compile (#167)

authored and
xinhe3
committed
[SW-222366] Switch tests to lazy mode (#174)

tbohutyn
authored and
xinhe3
committed
[SW-222366] Move env default to init (#178)

tbohutyn
authored and
xinhe3
committed
[SW-223106] Temporary disable mixed precision test (#180)

ulivne
authored and
xinhe3
committed
fp8 aware gptq (hybrid gptq) (#154)

authored and
xinhe3
committed
[SW-222513] OSError: does not appear to have a file named generation_config.json (#181)

xin3he
authored and
xinhe3
committed
Revert "fp8 aware gptq (hybrid gptq) (#154)" (#184)

tszulist-hbn
authored and
xinhe3
committed
[SW-221589] AutoRound W4A8 Quantization and Loading (#110)

authored and
xinhe3
committed
[SW-214855] - Set scale attributes in INC to reduce graph recompilations (#162)

nirda7
authored and
xinhe3
committed
[SW-222220] Quantising Llama3.2. 11B/90B fails with GC error (#179)

Yu-Zhou
authored and
xinhe3
committed
[SW-216623] Restore patch module to original before convert (#185)

xin3he
and
xinhe3
committed
[SW-218081] move htcore.hpu_set_env() to confest (#146)

xin3he
and
xinhe3
committed
[SW-221588] Unify weights in multi-cards for HF/INC format (N->M) (#172)

xin3he
and
xinhe3
committed
[SW-221594]Re-quantize the Official DeepSeek FP8 Model (#187)

authored and
xinhe3
committed
[SW-218277]Add support for mixtral with expert parallelism (#177)

kwisniewski98
authored and
xinhe3
committed
Update PatchedVLLMKVCache for deepseek performance (#194)

authored and
xinhe3
committed
[SW-224836] disable test_mixed_precision_gptq_fp8_quant_only_nlp (#208)

authored and
xinhe3
committed
Fix `PatchedMoeMatmul` and Get `num_experts` from Module (#202)

authored and
xinhe3
committed

Commits on Apr 25, 2025

Commits on Apr 28, 2025

Commits on Apr 29, 2025

fix bug in previous UT
xinhe3
committed