Repository navigation
v0.1.609
feat(speech): FastConformer encoder graph on ggml with per-stage parity (#737)
speech::encoderbuilds the Canary 1B v2 encoder from the loaded
tensors: dw_striding subsampling, host-built relative position table,
32 blocks with rel-pos attention onmul_mat+soft_max_ext, a conv
module with a composed sigmoid GLU,ssm_convdepthwise kernel and
BatchNorm folded at construction, half-step feed-forwards.- The CPU backend gets an abort callback hook;
ffibindsssm_conv,
conv_2d_dw_direct,reluandprec_set_acc.ggml_prec_set_acc
is optional, so older releases still load for the other mediagen
models; only the encoder refuses to build without it. subsampled_len(0)is 0 instead of an underflow.- An ignored e2e test compares six dump points of the NeMo 2.7.3
reference on the F16 and Q8_0 files; bounds are 1.5x the measured
values plus a relative rms bound, recorded in the tolerance table. It
sharescanary_fixtureandenv_dirwith the other Canary tests.
Closes #716