|
|
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 |
|
|
|
054d28573a
|
my DAC dataset again managed to only have some utterances with only 8 of 9 RVQ levels, this fixes an oversight from it
|
2024-08-09 21:18:01 -05:00 |
|
|
|
2a1794c084
|
ughghghhhh
|
2024-08-09 21:15:01 -05:00 |
|
|
|
ed373957e2
|
maybe not
|
2024-08-09 11:38:08 -05:00 |
|
|
|
c658a7b440
|
make loss scaling opt-in rather than automatically determined (because it seems a DAC-based model really doesnt like loss scaling)
|
2024-08-09 10:51:36 -05:00 |
|
|
|
d04f6911b4
|
oops
|
2024-08-08 19:38:55 -05:00 |
|
|
|
0aa59e6f3f
|
uncommented block that writes the metadata on HDF5 creation
|
2024-08-08 19:21:29 -05:00 |
|
|
|
79a6781c9e
|
fix vall_e.data --action=hdf5 actually transcribing because past me completely forgot it tried to already put the transcribe/process dataset scripts inside the module before
|
2024-08-08 07:51:42 -05:00 |
|
|
|
949339a3fa
|
do not include SDPA attention if there's no available SDPA backends
|
2024-08-06 20:42:39 -05:00 |
|
|
|
613024ec0d
|
ugh
|
2024-08-06 20:35:15 -05:00 |
|
|
|
eac353cd0b
|
busy work and cleanup while I wait for 1TB of audio to quantize... again.
|
2024-08-06 20:23:33 -05:00 |
|
|
|
f284c7ea9c
|
do mixed-precision for AMP inside the compress function itself, because the loudness function gripes when using a float16 (non-power of 2 lengths) or bfloat16 (something about views for bfloat16)
|
2024-08-06 15:08:37 -05:00 |
|
|
|
b6ba2cc8e7
|
tweaked vall_e.emb.process to instead process audio one file at a time instead of all the files for a given speaker to avoid OOMing on less-memory-filled systems with --low-memory
|
2024-08-06 14:24:40 -05:00 |
|
|
|
9710b06b74
|
tweaks and things
|
2024-08-06 08:17:25 -05:00 |
|
|
|
8bac8fe902
|
oops
|
2024-08-05 20:38:29 -05:00 |
|
|
|
134dac8c2b
|
re-adapted process_libritts.py to a 'better' way (better because it processed without needing to shuffle a bunch of things and adapt to cope or something)
|
2024-08-05 20:34:58 -05:00 |
|
|
|
3f73fcca29
|
oops
|
2024-08-05 20:12:13 -05:00 |
|
|
|
597441e48b
|
moved transcribe and process dataset scripts to vall_e/emb within the module itself, argparse-ified transcription script
|
2024-08-05 19:40:50 -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 |
|
|
|
10aaf840e7
|
added export option to convert Llama to MixtralMoE for another dumb experiment
|
2024-08-04 20:25:06 -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 |
|
|
|
d19f93a2c0
|
documentation update
|
2024-08-04 00:14:49 -05:00 |
|
|
|
2cb465018b
|
implicitly load either normal pickled weights or safetensors on loading the model
|
2024-08-03 23:34:18 -05:00 |
|
|
|
c09133d00f
|
added safetensors support (with metadata) and feed whatever torch.load/torch.save into it
|
2024-08-03 23:15:20 -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 |
|
|
|
ab673e0426
|
add cap for NAR-len training, to avoid any weird cases in early training where it'll just mess up and generate long lengths
|
2024-08-03 21:00:32 -05:00 |
|
|
|
4d2b88b164
|
throw exception if training, but no model is set to train (because i ran into this wondering what the hell was happening)
|
2024-08-03 20:51:23 -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 |
|
|
|
4456d3172b
|
that's what I get for testing without hdf5 on my previous machine....
|
2024-08-02 20:44:01 -05:00 |
|
|
|
7a77978096
|
oversight with using resize_modules
|
2024-08-02 20:28:49 -05:00 |
|
|
|
808a79ebaf
|
oops
|
2024-08-01 22:56:04 -05:00 |
|
|
|
443422ecb5
|
ugh, finally got some form of offloading working (need to test if it works on different GPUs, but GPU and CPU offloading seems to work in the test trainer)
|
2024-08-01 22:43:39 -05:00 |
|
|
|
c9ec6b28ef
|
it actually wasn't working because Engines.__init__() automatically moves the entire module to the requested device, which was being called after offloading the model in the test trainer (and it seems I cant do it without injecting a bunch of shit in modeling_llama.py)
|
2024-08-01 20:56:28 -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 |
|
|
|
52d13b321f
|
I rather have it default to non-strict loading instead so I can clean up YAMLs
|
2024-07-30 22:24:38 -05:00 |
|
|
|
d7c6be6f78
|
fix weird regression in handling checkpoints when backend is local, but deepspeed checkpoints are in (it was handled with LoRA loading but not real loading...)
|
2024-07-30 22:15:56 -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 |
|
