Fix accuracy issues with Gemma models - #1448

Merged

kunal-vaishnavi merged 29 commits into

mainfrom

kvaishnavi/gemma3-mm

May 9, 2025

Commits on Apr 15, 2025

Initial implementation of bfloat16 support
kunal-vaishnavi
committed
cd5fc2c

Commits on Apr 21, 2025

Refactor multi-cache calculation
kunal-vaishnavi
committed
5fa15f2
Fix how caches are saved in if node
kunal-vaishnavi
committed
084936a

Commits on Apr 24, 2025

Commits on Apr 25, 2025

Merge branch 'kvaishnavi/bf16' into kvaishnavi/gemma-mixed-precision
kunal-vaishnavi
committed
e82b240
Generate Gemma-3 4B text-only model
kunal-vaishnavi
committed
7d2fc55
Cast MatMul from FP32 to BF16
kunal-vaishnavi
committed
89040f6

Commits on Apr 28, 2025

Merge branch 'main' into kvaishnavi/gemma3-mm
kunal-vaishnavi
committed
795d20b
Add missing name for constant node
kunal-vaishnavi
committed
c24d40d

Commits on Apr 29, 2025

GQA bf16 support (#1429)
Nenad Banfic (nenad1002)
authored
b1ec3bd
Fast Gelu
Nenad Banfic (nenad1002)
committed
1e541b6

Commits on May 2, 2025

Add Gemma3 bf16 related changes (#1441)
Nenad Banfic (nenad1002)
authored
be4e533

Commits on May 7, 2025

Commits on May 8, 2025

Refactor layernorm
Nenad Banfic (nenad1002)
committed
ed9eda0
Make sure bias is not used before being defined
Nenad Banfic (nenad1002)
committed
30077d0
Redesign and refactor how casting is done
kunal-vaishnavi
committed
577e4b3
Fix embedding check to be more generic
kunal-vaishnavi
committed
a6bb465
Add missing cast name
kunal-vaishnavi
committed
01b2a49
Revert quantization import consolidation until CIs are updated
kunal-vaishnavi
committed
b6d8b94
Remove unneeded parentheses
kunal-vaishnavi
committed
3eb173d
Add more granular control over LayerNorm casts
kunal-vaishnavi
committed
f8f6e02

Commits on May 9, 2025

Use same chat template for Gemma models
kunal-vaishnavi
committed
cdd273c