model: GraniteSpeech5ForCTC (Turbo CTC) #29446
gabe-l-hart
wants to merge
36 commits into
ggml-org:master
from
gabe-l-hart:GraniteSpeechCTC
feat: Add constants, tensor_mapping, and gguf_writer helpers for Gran…
35bf7444
feat: Add option to skip n_embd_text check for MMPROJ models without …
2cbd789f
feat: Add chkhsh exception for granite-speech-5
e23a7d98
feat: Add encoder and mmproj FE conversion for granite-speech-5
ebc12ed0
feat(api): Add public llama_n_outputs
5210decd
feat: Automatically raise ubatch size for encoder-only models to fit …
1e4f98f0
feat: Compute encoder output size explicitly to allow for time downsa…
7988022c
feat: Make subsample_layers a per-layer factor rather than an indicator
d13fe6d8
feat: Add llama_hparams handling for per-layer subsample_factor
0451f2d3
feat: Explicitly map pretokenizer to default
20326b6e
feat: Add CTC encoder hparams block
6ef1a902
feat: Add c++ constants for arch, GGUF kvs, and tensors
ffc5b134
feat: Add llama_model tensor pointers
4cc1b67b
feat: Add granite-speech-5 model impl
ead6674e
fix: Typo in granite_speech-5
3b6117a0
fix: Remove unused python-side hparam constants
43ae45d9
fix: Use native output_dim instead of counted n_vocab for output tensors
afcb8291
feat: Add interface in common for running ctc forward pass and decoding
56fdcd3d
refactor(convert): granite-speech-5 -> granite_speech_5
5748fc83
refactor(c++): granite-speech-5 -> granite_speech_5
716df2df
feat: Add transformer-less frontend clip model for granite-speech-5
983a44dc
feat: Add clip model and hparam plumbing for granite-speech-5 frontend
286fb5bb
feat: Add granite-speech-5 preprocessor
3a537720
feat(server): Add server task and response type for transcribe
d01cbc61
feat(server): Wire through encoder-only transcriptions handling
511da47c
fix: Use the actual unpadded number of tokens for output
f37a0409
fix(convert): Don't insert an extra blank in the vocabulary
6a92302a
Revert "fix: Use native output_dim instead of counted n_vocab for out…
1b843565
fix: Always use llama_n_tokens in common_ctc_greedy_decode
1e497b50
Revert "fix: Use the actual unpadded number of tokens for output"
6131ff97
feat: Add CTC encoder params to llama-model-saver
3e2cbe39
test: Add CTC params for G5 speech in test-llama-archs
023e0345
fix: Serialize new tensors for CTC out_mid in llama-model-saver
6aeca0ab
test: Skip g-speech-5 for tests that run tokens through
f0259423
fix: Allow non-sampling models through logits memory-free check after…
08d6392c
fix: Fix bad HF link in base.py
61563269
Login to write a write a comment.
Login via GitHub