HGQ v2 support (part 1)#1154

Closed

calad0i wants to merge 69 commits intofastmachinelearning:mainfrom

calad0i:s-quark-pr

Commits on Feb 6, 2025

Add hint on import failure
calad0i
committed
import converter dependencies lazily
calad0i
committed
keras v3 object based parser
calad0i
committed
sequential and i/o tensor name parsing fix
calad0i
committed
support activation layers
calad0i
committed
consistent v2 weight reader behavior
calad0i
committed
add v3 conv handlers
calad0i
committed
add test
calad0i
committed
pre-commit fix
calad0i
committed
revert keras v2 converter
calad0i
committed
make reshape handler compatiable with keras v3
calad0i
committed
general einsum support for io_parallel and latency
calad0i
committed
add tests for einsumdense
calad0i
committed
keras v3 converter clean-up
calad0i
committed
add symbolic quantized interval
calad0i
committed
preliminary bit-exact precision derivation opt pass
calad0i
committed
squark layer support start
calad0i
committed
fix einsum_dense precision computation
calad0i
committed
add leftover
calad0i
committed
qdense fix
calad0i
committed
support batch_norm
calad0i
committed
support merge layers
calad0i
committed
support bit-exact q_einsum and fix precision trace for multi inp layers
calad0i
committed
add einsum test
calad0i
committed
declare all softmax attrs in layer class
calad0i
committed
fix lazy import in handler
calad0i
committed
cleanup einsum handler
calad0i
committed
cleanup einsum handler
calad0i
committed
more granular control over softmax for vivado
calad0i
committed
properly propagate inv/exp_table_size
calad0i
committed
support bit-exact softmax for stable impl
calad0i
committed
bit-exact softmax fix and leftovers
calad0i
committed
softmax table fixer update
calad0i
committed
support input scaler in softmax
calad0i
committed
support multidim parallel softmax
calad0i
committed
fuse quantizer when possible
calad0i
committed
partial activation, fix input precision in SAT mode
calad0i
committed
fix padded convXd precition derivation rule
calad0i
committed
add unary lut support
calad0i
committed
fix bit-exact corner case introduced by reverse flow
calad0i
committed
general data_t inference
calad0i
committed
softmax compatbility
calad0i
committed
fix typo in einsum handler
calad0i
committed
fix more typos
calad0i
committed
MHA :tada:
calad0i
committed
fix einsum and softmax template typos
calad0i
committed
assert einsum ops doesnot include direct sum operation
calad0i
committed
style
calad0i
committed
fix mha layer indexing
calad0i
committed
switch to model opt
calad0i
committed
pooling layers
calad0i
committed
handle stray inputs
calad0i
committed
fix pooling layer accum_t
calad0i
committed
bit-exact concatenate
calad0i
committed
add comments
calad0i
committed
skip non-bit-exact compatiable softmax in bit-exact pass
calad0i
committed
fix activation matching, fix bw edge case handle
calad0i
committed
fixes in bw inference
calad0i
committed
warn double quantizer
calad0i
committed
fix softmax bw and table_size in non-bit-exact case
calad0i
committed
fix dependency
calad0i
committed
formatting after rebase
calad0i
committed
add kif cache
calad0i
committed
fix edge cases, allow fixedquantizer bw shrink when possible
calad0i
committed
fix global pooling bw inference
calad0i
committed
rm linear layer if FixedQuantizer presents
calad0i
committed
add parallelization factor support for squaark QDense
calad0i
committed
warn at bw inference overflow, instead of crash
calad0i
committed
minor fix on weight type var detection
calad0i
committed