|
|
42fafbaaca
|
actually fixed knowledge distillation because of errant -inf logits causing problems and needed to be filtered (and splitting text language / output audio language because it helps)
|
2024-12-06 21:55:20 -06:00 |
|
|
|
23d402bf01
|
added knowledge distillation in the trainer (sadly it is not agnostic because of the grave mistake of further processing the batch within the forward pass, so subsequent calls do not match......)
|
2024-12-05 23:05:52 -06:00 |
|
|
|
84a05acb6d
|
touch ups in docs
|
2024-12-02 19:10:42 -06:00 |
|
|
|
4aa685e749
|
what has science done
|
2024-11-22 16:45:40 -06:00 |
|
|
|
147219a5e0
|
huge oversight in the attention masking......... (i realized I have not been providing a non-causal mask to non-causal tasks)
|
2024-11-22 13:44:43 -06:00 |
|
|
|
8aafae91fd
|
dont use timeembedding
|
2024-11-21 23:14:52 -06:00 |
|
|
|
2cef97e43f
|
cleanup
|
2024-11-21 23:08:43 -06:00 |
|
|
|
190a917b3e
|
I did it.
|
2024-11-19 12:24:33 -06:00 |
|
|
|
6cfdf94bf9
|
swap priority to use nar-len if available, added notes
|
2024-11-18 09:40:04 -06:00 |
|
|
|
069b27570f
|
set option to set training masking ratio (I don't think for tts a fixed masking ratio is beneficial since the magic of the AR+NAR is being able to still reference the prior sequence of tokens for predicting things)
|
2024-11-17 17:04:07 -06:00 |
|
|
|
88d840218d
|
default set cfg strength to 3.0 since the reference model is updated
|
2024-11-17 10:23:40 -06:00 |
|
|
|
a3e1fa3518
|
ugh
|
2024-11-17 09:28:33 -06:00 |
|
|
|
23fdba0c98
|
tweaks and changes
|
2024-11-16 15:49:06 -06:00 |
|
|
|
2fbeacfe92
|
ugh
|
2024-11-14 22:18:33 -06:00 |
|
|
|
39096f8ff3
|
redid loss calculation to be cleaner, and position ID generation, and other things (I might need to train the NAR-len from scratch and not resume from an existing checkpoint.........)
|
2024-11-14 22:17:47 -06:00 |
|
|
|
e412e98125
|
ugh
|
2024-11-14 07:34:22 -06:00 |
|
|
|
c00fc18b62
|
actually use the right embedding for nar-len
|
2024-11-13 18:04:04 -06:00 |
|
|
|
3ea8a610d6
|
fix STT
|
2024-11-13 14:27:15 -06:00 |
|
|
|
910033343c
|
overhauled how the right resp level / classifier gets picked to avoid cringemath
|
2024-11-13 13:31:17 -06:00 |
|
|
|
269648605e
|
move NAR-len rvq level 0 to separate embedding
|
2024-11-13 11:38:58 -06:00 |
|
|
|
be83ddabaa
|
better causal-ness for split loss calc, and also do masking for NAR-len for it
|
2024-11-13 10:17:52 -06:00 |
|
|
|
6b76419123
|
ugh
|
2024-11-13 09:54:20 -06:00 |
|
|
|
ad7cfffc00
|
NAR-len RVQ-0 was being trained causally.............
|
2024-11-13 09:43:50 -06:00 |
|
|
|
8286aa54c8
|
do not pass timestep token/embedding since it doesn't seem to matter at all after all, fixed training masking rate to 80% because a paper said so
|
2024-11-13 09:07:10 -06:00 |
|
|
|
0f2584eba7
|
new meme sampler PogChamp new meme sampler PogChamp (it sort of helps?)
|
2024-11-12 22:30:09 -06:00 |
|
|
|
663f07038d
|
haha... (do not create a token dropout/noise mask when not training (this sadly didnt fix NAR-len output))
|
2024-11-12 16:41:58 -06:00 |
|
|
|
8927bad7bc
|
actually fixed rep pen (for ar and nar, it seems to help with nar unmasking)
|
2024-11-11 21:40:19 -06:00 |
|
|
|
2f56696506
|
overhauled inference/sampler kwargs to stop being a bloated mess
|
2024-11-11 20:21:16 -06:00 |
|
|
|
9cb0b6901b
|
unified nar.py into ar_nar.py
|
2024-11-10 12:19:48 -06:00 |
|
|
|
a9d2faf2d7
|
all I can do now until I wait for the model to (re)train for pure NAR
|
2024-11-09 22:57:34 -06:00 |
|
|
|
ad7e290a5e
|
ugh (ROCm seems to silently clamp any token value >= logits.shape[-1] for loss calculation, while cuda will throw an assert, making it hard to find this dumb fuckup)
|
2024-11-09 19:40:02 -06:00 |
|
|
|
943fe70c10
|
I don't know why this fixes an assert thrown but it does
|
2024-11-09 19:04:13 -06:00 |
|
|
|
f50d92ba6c
|
Almost made a mistake
|
2024-11-09 18:12:54 -06:00 |
|
|
|
c6a38693a2
|
This better work
|
2024-11-09 18:04:59 -06:00 |
|
|
|
8b3d1cf70a
|
Something's Wrong
|
2024-11-09 15:07:43 -06:00 |
|
|
|
69b0b3b854
|
set timestep tensor to whatever the time embedding's dtype is because it'll gripe under amp
|
2024-11-09 00:11:16 -06:00 |
|
|
|
5a09a5f6e9
|
I forgot about the time embedding...
|
2024-11-08 22:46:26 -06:00 |
|
|
|
811b15d280
|
I suppose I just have a shit training method since the sampler is as solid as I can get it...............
|
2024-11-08 22:05:41 -06:00 |
|
|
|
13b54953bd
|
agony
|
2024-11-08 13:34:39 -06:00 |
|
|
|
c127c4e488
|
'borrowed' a sampling scheduler for NAR-len's RVQ level 0 (better than before, but still not good enough)
|
2024-11-07 21:19:14 -06:00 |
|
|
|
e108c54daf
|
new NAR-len training paradigm......
|
2024-11-07 11:32:11 -06:00 |
|
|
|
ed174c589e
|
ugh
|
2024-11-07 09:19:21 -06:00 |
|
|
|
5698188824
|
あたしって、ほんとバカ
|
2024-11-07 09:10:18 -06:00 |
|
|
|
105ed51159
|
I guess I'll fall for the NAR-len meme again (I don't know where my previous weights are, so I need to train it again to test something)
|
2024-11-06 19:17:12 -06:00 |
|
|
|
9e65e05e83
|
more windows specific fixes, limit gradio to <5.0.0 on linux (it works on windows, but not on my linux machine tm)
|
2024-11-04 18:00:33 -06:00 |
|
|
|
d229725c76
|
more adjustments (adjustments of early-exit entropy/varentropy thresholds, default rep pen being 1.5, experimental refine-on-stop, etc.)
|
2024-11-03 18:31:28 -06:00 |
|
|
|
aee08b7307
|
changed layerskip float16 training warning (since it didnt seem to fry on my 4xV100 system)
|
2024-11-03 09:58:29 -06:00 |
|
|
|
3826f9bae4
|
saner mask creation? (it doesnt matter, kv cache wont work)
|
2024-11-02 21:00:21 -05:00 |
|
|
|
ded746e157
|
very, very naive layerskip speculative sampling (it just checks if the current layer's state is good enough)
|
2024-11-02 11:49:05 -05:00 |
|
|
|
ec79230965
|
shuffled web UI options hidden by cfg.experimental to its own tab, expose early exit selection to inferencing (it kinda works naively, still need to implement self-speculation)
|
2024-11-01 21:30:06 -05:00 |
|
|
|
9b6c57bc57
|
third time's the charm (for some reason it escaped me that I should treat early exit loss as an aux_loss to be used with the normal loss, as if I was training a MoE's router)
|
2024-11-01 12:50:37 -05:00 |
|
|
|
76ebef45dc
|
off-by-one...
|
2024-10-31 13:24:48 -05:00 |
|
|
|
b63293cbbe
|
ugh
|
2024-10-30 22:49:11 -05:00 |
|
|
|
a22534e8f4
|
layer skip training implemented (need to gut the inferencing from the repo, and to actually see if the model can benefit from this)
|
2024-10-30 20:05:45 -05:00 |
|
|
|
8eb9a4056b
|
modified default arguments (ar temp = 0 and rep pen = 1.125 seems to be stable, at least given the few things i tested), do not pass top k/top p/min p to NAR even though technically none of those things should matter when greedy sampling
|
2024-10-22 18:12:39 -05:00 |
|
|
|
fc8dfd8617
|
made greedy AR sampling viable (and preferable), with caveats (per comment in vall_e.models.ar_nar)
|
2024-10-18 16:55:00 -05:00 |
|
|
|
84005c5b00
|
entropix apparently processes the entire sequence of logits but it falls apart when doing that
|
2024-10-13 12:01:12 -05:00 |
|
|
|
c800d28bb8
|
respect attention defined in the yaml for web UI (which might explain why theres been a discrepancy in outputs for me)
|
2024-10-13 11:02:24 -05:00 |
|
|
|
d405f243d4
|
at wits end in trying to output the right attention scores
|
2024-10-12 23:53:13 -05:00 |
|
|
|
04e983b86b
|
modified demo page to be more modular with demoing comparisons, actually provide a path to use modified naive attention, entropix sampling is not tied to an experimental yaml flag now
|
2024-10-12 11:27:55 -05:00 |
|
|
|
666e8038fb
|
ugh
|
2024-10-12 10:41:35 -05:00 |
|
|
|
d6f7c86a5c
|
entropix tweaks (it doesn't output garbage but it loves to go for silence)
|
2024-10-12 09:46:18 -05:00 |
|
|
|
d0ab7d755a
|
added min-p (really does not seem useful since it's very sensitive), more tweaks to entropix
|
2024-10-11 22:36:06 -05:00 |
|
|
|
bef43a0c18
|
added experimental entropix sampling support
|
2024-10-11 21:18:26 -05:00 |
|
|
|
acdce66d4e
|
readme tweaks, set the (unused) default model download URL back to the base ar+nar-llama-8 model, as ar+nar-tts+stt-llama-8 was renamed back to it since it performs well
|
2024-10-05 22:53:53 -05:00 |
|
|
|
84c7419001
|
faster
|
2024-10-04 22:30:47 -05:00 |
|
|
|
a507b769a1
|
sped up inferencing by not doing .tolist() for rep pen / length pen (and a bug fix in the web UI from prev commit)
|
2024-10-04 22:18:20 -05:00 |
|
|
|
54203c059d
|
validated rep pen for STT (sometimes needed to wrangle the model)
|
2024-09-08 08:30:30 -05:00 |
|
|
|
6a967f91b9
|
oops
|
2024-09-07 22:13:49 -05:00 |
|
|
|
4bd9bb39c8
|
webui for STT (still need to bake the model to handle it better, a few hours so far has it generate what looks like a normal transcription but does not correlate to the audio right now)
|
2024-09-06 15:13:04 -05:00 |
|
|
|
341e19162b
|
fixes, again
|
2024-09-06 11:41:41 -05:00 |
|
|
|
413097f5f7
|
fixes
|
2024-09-05 21:42:59 -05:00 |
|
|
|
54547b74d8
|
experimental implementation of STT (need to actually test on a model, test trainer seems to work)
|
2024-09-05 20:43:20 -05:00 |
|
|
|
b7b99a25f1
|
added ability to specify attention backend for CLI and webui (because im tired of editing the yaml)
|
2024-08-26 19:33:51 -05:00 |
|
|
|
0d706ec6a1
|
added fused_attn (triton-based fused attention) and simply just query for flash_attn under rocm
|
2024-08-26 19:13:34 -05:00 |
|
|
|
6b0891448c
|
pain (some shit to try and get some flash attention for ROCm (gfx1100) through triton fused attention but no good)
|
2024-08-25 20:07:27 -05:00 |
|
|
|
40e1799adc
|
fixed xformers and flash_attn to actually work now
|
2024-08-19 01:03:35 -05:00 |
|
|
|
29c35528e5
|
the sooner I accept there's no FA for V100s the sooner I'll go to bed
|
2024-08-18 23:54:33 -05:00 |
|
|
|
d636edd3a2
|
added flash_attn LlamaAttention (including flash_attn==1.0.9)
|
2024-08-18 20:51:14 -05:00 |
|
|
|
2a1794c084
|
ughghghhhh
|
2024-08-09 21:15:01 -05:00 |
|
|
|
d04f6911b4
|
oops
|
2024-08-08 19:38:55 -05:00 |
|
|
|
949339a3fa
|
do not include SDPA attention if there's no available SDPA backends
|
2024-08-06 20:42:39 -05:00 |
|
|
|
7cdfa3dc0c
|
updated process_datasets.py, added argparsing so I can mostly stop manually editing things, and some other cleanup
|
2024-08-05 15:59:25 -05:00 |
|
|
|
debcc93e7e
|
add adapted MixtralAttention for when I make a bad decision to actually train a MoE
|
2024-08-04 22:03:22 -05:00 |
|
|
|
3a65cc4b22
|
fix issue with sft and shared tensors...
|
2024-08-04 19:56:21 -05:00 |
|
|
|
23f3b56fda
|
oops
|
2024-08-04 08:18:57 -05:00 |
|
|
|
6a733eb2ed
|
changed torch.Tensor().to(device, dtype) to just torch.tensor(..., device, dtype) because it's been bothering my autism that I'm creating tensors then converting rather than creating with the right device/dtype, some 'optimization' to compile the model but it doesnt seem to do anything useful
|
2024-08-03 22:10:21 -05:00 |
|
|
|
d0a5c7eca2
|
more coping with the NAR len
|
2024-08-03 20:23:36 -05:00 |
|
|
|
11fa3da665
|
some cleanup, fixed the wrapper attention to explicitly use other sdpa backends
|
2024-08-03 19:51:00 -05:00 |
|
|
|
9564ecda43
|
wrapper attention class for other sdpa backends + xformers seems to have broke...
|
2024-08-03 15:12:11 -05:00 |
|
|
|
9e1989be1b
|
tweaked initial NAR pass's initial token embeddings to use a different value, or osmething
|
2024-08-03 09:01:37 -05:00 |
|
|
|
26f74c5739
|
somehow fixed non-unified position IDs for the NAR-len
|
2024-08-03 08:43:42 -05:00 |
|
|
|
66407e5bdb
|
tweaks for the NAR-len model, maybe
|
2024-08-03 08:40:39 -05:00 |
|
|
|
97c5241bef
|
fixes, throw an exception when using NAR only model with non-unified position IDs, since for some reason it outputs garbage for the NAR
|
2024-08-02 22:25:49 -05:00 |
|
|
|
b4c895114c
|
naive model offloading support (handles automatically splitting parts of the model to requested device per memory constraints, either inferred or requested in the yaml, input tensors are automatically migrated to the right device, it SEEMS to work for training under the test trainer when split between GPU and CPU) (this was specifically only because that Flux imagegen model released so I can test it there)
|
2024-08-01 20:12:06 -05:00 |
|
|
|
387358bc8a
|
fixes for the NAR-len model, and documentation some config options, and a better way to handle resizing modules on state_dict load
|
2024-07-31 20:35:09 -05:00 |
|
|
|
07f8e2ad06
|
added option to set the causal size (how many tokens to sample per AR step), but requires the model to be trained for this (which explains why recurrent chunk sampling just doesn't work for the retnet tests, obvious in hindsight)
|
2024-07-30 20:53:51 -05:00 |
|
|
|
ebf848d249
|
possible speedup for samplers that require a list of previous tokens (the DRY sampler made me realize that I should copy the tolist() thing from the rep pen sampler for everything else)
|
2024-07-29 20:23:26 -05:00 |
|
|
|
55b0121b1a
|
trying (and failing) to nail a weird regression in fancier attentions
|
2024-07-29 19:53:37 -05:00 |
|
|
|
c2f5b916fc
|
added what I think is DRY sampling
|
2024-07-29 19:15:07 -05:00 |
|
