Problem with TTS in 2.8

[Jasonthefirst]

We are using LocalAI in Docker but have Problems with all TTS models described in TTS in LocalAI .

But when calling the following curl:

curl http://localhost:8080/tts -H "Content-Type: application/json" -d '{ "backend": "bark", "input":"Hello!" }' | aplay

we get the following error:

stderr OSError: /opt/conda/envs/transformers/lib/python3.11/site-packages/torchaudio/lib/libtorchaudio.so: undefined symbol: _ZN2at4_ops10zeros_like4callERKNS_6TensorEN3c108optionalINS5_10ScalarTypeEEENS6_INS5_6LayoutEEENS6_INS5_6DeviceEEENS6_IbEENS6_INS5_12MemoryFormatEEE

This error is thrown with bark, qoqui and Vall-E-X. Piper works.

LocalAI version:
v2.8.0-cublas-cuda12-ffmpeg

Environment, CPU architecture, OS, and Version:
Linux aifb-bis-mlpc 5.15.0-92-generic #102-Ubuntu SMP Wed Jan 10 09:33:48 UTC 2024 x86_64 x86_64 x86_64 GNU/Linux

To Reproduce
Run v2.8.0-cublas-cuda12-ffmpeg LocalAI on Server an the curl command.

Expected behavior
LocalAI shouldn't return an error but a tts file.

Logs
I added a log file. _Shared_LocalAI_logs.txt

[mudler]

this should be already fixed in master images - @Jasonthefirst could you please test it out?

[golgeek]

@mudler not sure if your fix was also deployed for vllm, but I just tried loading a model with vllm backend with the master image and got the same error:

3:11PM DBG GRPC(casperhansen/mixtral-instruct-awq-127.0.0.1:45981): stderr Traceback (most recent call last):
3:11PM DBG GRPC(casperhansen/mixtral-instruct-awq-127.0.0.1:45981): stderr   File "/build/backend/python/vllm/backend_vllm.py", line 13, in <module>
3:11PM DBG GRPC(casperhansen/mixtral-instruct-awq-127.0.0.1:45981): stderr     from vllm import LLM, SamplingParams
3:11PM DBG GRPC(casperhansen/mixtral-instruct-awq-127.0.0.1:45981): stderr   File "/opt/conda/envs/transformers/lib/python3.11/site-packages/vllm/__init__.py", line 3, in <module>
3:11PM DBG GRPC(casperhansen/mixtral-instruct-awq-127.0.0.1:45981): stderr     from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
3:11PM DBG GRPC(casperhansen/mixtral-instruct-awq-127.0.0.1:45981): stderr   File "/opt/conda/envs/transformers/lib/python3.11/site-packages/vllm/engine/arg_utils.py", line 6, in <module>
3:11PM DBG GRPC(casperhansen/mixtral-instruct-awq-127.0.0.1:45981): stderr     from vllm.config import (CacheConfig, ModelConfig, ParallelConfig,
3:11PM DBG GRPC(casperhansen/mixtral-instruct-awq-127.0.0.1:45981): stderr   File "/opt/conda/envs/transformers/lib/python3.11/site-packages/vllm/config.py", line 9, in <module>
3:11PM DBG GRPC(casperhansen/mixtral-instruct-awq-127.0.0.1:45981): stderr     from vllm.utils import get_cpu_memory, is_hip
3:11PM DBG GRPC(casperhansen/mixtral-instruct-awq-127.0.0.1:45981): stderr   File "/opt/conda/envs/transformers/lib/python3.11/site-packages/vllm/utils.py", line 11, in <module>
3:11PM DBG GRPC(casperhansen/mixtral-instruct-awq-127.0.0.1:45981): stderr     from vllm._C import cuda_utils
3:11PM DBG GRPC(casperhansen/mixtral-instruct-awq-127.0.0.1:45981): stderr ImportError: /opt/conda/envs/transformers/lib/python3.11/site-packages/vllm/_C.cpython-311-x86_64-linux-gnu.so: undefined symbol: _ZN2at4_ops15to_dtype_layout4callERKNS_6TensorEN3c108optionalINS5_10ScalarTypeEEENS6_INS5_6LayoutEEENS6_INS5_6DeviceEEENS6_IbEEbbNS6_INS5_12MemoryFormatEEE

[mudler]

@golgeek I've tried only with TTS models (vall-e-x specifically), can you confirm that? please open up another issue for vLLM

[Jasonthefirst]

We tested it with the master branch (master-cublas-cuda12-ffmpeg). As input we used the standard curl and we got the following error:

curl:

curl
http://localhost:8080/tts
-H "Content-Type: application/json" -d '{
   "backend": "bark",
   "input":"Hello!"
}' | aplay

4:32PM DBG Loading model in memory from file: /models/model_configuration
4:32PM DBG Loading Model  with gRPC (file: /models/model_configuration) (backend: hello!): {backendString:Hello! model: threads:0 assetDir:/tmp/localai/backend_data context:{emptyCtx:{}} gRPCOptions:0xc000230800 externalBackends:map[autogptq:/build/backend/python/autogptq/run.sh bark:/build/backend/python/bark/run.sh coqui:/build/backend/python/coqui/run.sh diffusers:/build/backend/python/diffusers/run.sh exllama:/build/backend/python/exllama/run.sh exllama2:/build/backend/python/exllama2/run.sh huggingface-embeddings:/build/backend/python/sentencetransformers/run.sh mamba:/build/backend/python/mamba/run.sh petals:/build/backend/python/petals/run.sh sentencetransformers:/build/backend/python/sentencetransformers/run.sh transformers:/build/backend/python/transformers/run.sh transformers-musicgen:/build/backend/python/transformers-musicgen/run.sh vall-e-x:/build/backend/python/vall-e-x/run.sh vllm:/build/backend/python/vllm/run.sh] grpcAttempts:20 grpcAttemptsDelay:2 singleActiveBackend:true parallelRequests:true}
[172.24.0.1]:37026 500 - POST /tts

[golgeek]

@golgeek I've tried only with TTS models (vall-e-x specifically), can you confirm that? please open up another issue for vLLM

Sorry, the error seemed too suspiciously similar, and I thought it might be the same origin.

Was coming back to report the same as @Jasonthefirst.

And I opened #1710 for vLLM.

[mudler]

@golgeek / @Jasonthefirst any chance you can give #1711 a shot?

[mudler]

waiting for feedback, merged the PR so master images are going to be built soon so we can try it out much easier just consuming master images

[golgeek]

Sorry that it took me forever to realize that the images weren't pushed, then an equal amount of time to build a docker image from your branch.

I just ran a quick test for vLLM and the model loaded successfully, so I'd say it's fixed but maybe it's better to wait a bit more and confirm with the images from the master branch.

[golgeek]

I ran some tests again with master images, and can confirm that #1710 is fixed (just closed the issue, thanks a lot!).

As for this issue specifically, I tested with master-cublas-cuda12-ffmpeg (sha256:de26b09328fea0bd57ff2e14ae28ba9a54ca489a1bc96208131f8d4c1d494672), and while the initial error is definitely fixed, there still seems to be something wrong as, when curling:

curl http://localhost:8080/tts -H "Content-Type: application/json" -d '{"backend": "bark","input":"Hello!"}'
{"error":{"code":500,"message":"grpc process not found: /tmp/localai/backend_data/backend-assets/grpc/hello!. some backends(stablediffusion, tts) require LocalAI compiled with GO_TAGS","type":""}}

I'm getting:

3:03PM DBG Request for model:
3:03PM INF Loading model with backend Hello!
3:03PM DBG Loading model in memory from file: /build/models
3:03PM DBG Loading Model  with gRPC (file: /build/models) (backend: hello!): {backendString:Hello! model: threads:0 assetDir:/tmp/localai/backend_data context:{emptyCtx:{}} gRPCOptions:0xc0001f6000 externalBackends:map[autogptq:/build/backend/python/autogptq/run.sh bark:/build/backend/python/bark/run.sh coqui:/build/backend/python/coqui/run.sh diffusers:/build/backend/python/diffusers/run.sh exllama:/build/backend/python/exllama/run.sh exllama2:/build/backend/python/exllama2/run.sh huggingface-embeddings:/build/backend/python/sentencetransformers/run.sh mamba:/build/backend/python/mamba/run.sh petals:/build/backend/python/petals/run.sh sentencetransformers:/build/backend/python/sentencetransformers/run.sh transformers:/build/backend/python/transformers/run.sh transformers-musicgen:/build/backend/python/transformers-musicgen/run.sh vall-e-x:/build/backend/python/vall-e-x/run.sh vllm:/build/backend/python/vllm/run.sh] grpcAttempts:20 grpcAttemptsDelay:2 singleActiveBackend:false parallelRequests:false}

It appears there might be a mixup between the backend and input fields as LocalAI tries to load the Hello! backend, though the TTSEndpoint code looks totally legit.

[mudler]

I ran some tests again with master images, and can confirm that #1710 is fixed (just closed the issue, thanks a lot!).

As for this issue specifically, I tested with master-cublas-cuda12-ffmpeg (sha256:de26b09328fea0bd57ff2e14ae28ba9a54ca489a1bc96208131f8d4c1d494672), and while the initial error is definitely fixed, there still seems to be something wrong as, when curling:

curl http://localhost:8080/tts -H "Content-Type: application/json" -d '{"backend": "bark","input":"Hello!"}'
{"error":{"code":500,"message":"grpc process not found: /tmp/localai/backend_data/backend-assets/grpc/hello!. some backends(stablediffusion, tts) require LocalAI compiled with GO_TAGS","type":""}}

I'm getting:

3:03PM DBG Request for model:
3:03PM INF Loading model with backend Hello!
3:03PM DBG Loading model in memory from file: /build/models
3:03PM DBG Loading Model  with gRPC (file: /build/models) (backend: hello!): {backendString:Hello! model: threads:0 assetDir:/tmp/localai/backend_data context:{emptyCtx:{}} gRPCOptions:0xc0001f6000 externalBackends:map[autogptq:/build/backend/python/autogptq/run.sh bark:/build/backend/python/bark/run.sh coqui:/build/backend/python/coqui/run.sh diffusers:/build/backend/python/diffusers/run.sh exllama:/build/backend/python/exllama/run.sh exllama2:/build/backend/python/exllama2/run.sh huggingface-embeddings:/build/backend/python/sentencetransformers/run.sh mamba:/build/backend/python/mamba/run.sh petals:/build/backend/python/petals/run.sh sentencetransformers:/build/backend/python/sentencetransformers/run.sh transformers:/build/backend/python/transformers/run.sh transformers-musicgen:/build/backend/python/transformers-musicgen/run.sh vall-e-x:/build/backend/python/vall-e-x/run.sh vllm:/build/backend/python/vllm/run.sh] grpcAttempts:20 grpcAttemptsDelay:2 singleActiveBackend:false parallelRequests:false}

It appears there might be a mixup between the backend and input fields as LocalAI tries to load the Hello! backend, though the TTSEndpoint code looks totally legit.

ouch, good catch, this is a regression introduced in #1692.