* chore(neuron): bump version to 0.2.0
* refactor(neuron): use named parameters in inputs helpers
This allows to hide the differences between the two backends in terms of
input parameters.
* refactor(neuron): remove obsolete code paths
* fix(neuron): use neuron_config whenever possible
* fix(neuron): use new cache import path
* fix(neuron): neuron config is not stored in config anymore
* fix(nxd): adapt model retrieval to new APIs
* fix(generator): emulate greedy in sampling parameters
When on-device sampling is enabled, we need to emulate the greedy
behaviour using top-k=1, top-p=1, temperature=1.
* test(neuron): update models and expectations
* feat(neuron): support on-device sampling
* fix(neuron): adapt entrypoint
* tests(neuron): remove obsolete models
* fix(neuron): adjust test expectations for llama on nxd
* forward and tokenize chooser use the same shape
concate or filter happened to cpu tensor to avoid dynamic shape in hpu
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* use hpu set seed
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
---------
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* clean cuda/rocm code in hpu backend, enable flat_hpu
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix TP in pageattn
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* adjust block table in hpu to improve performance
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* enable all the model. not testet yet
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* use tensor cache in hpu graph to avoid replay issue
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* add moe support, fix qwen/mistral/mixtral crash
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix phimoe issue
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* gpt_bigcode could also go pageattn
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* enable dbrx remove some unused code
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* multi-modality initial PR
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* adjust warmup and enable vlm
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix incorrect output in qwen2 idefics if hpu graph is used
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* remove unused quantization code and enable awq/gptq int4
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix gptq issue
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* enable fp8
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* warmup prefill
remove model where pageattn is not used, set block table to None since it's not used
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* add warmup_decode
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* warmup decode
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* remove block_tables and prefill_cache_indices which will lead to dynamic shape
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix comment
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* missing gptj change...
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix some issue
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* remove torch.where to fix incorrect output in hpu graph model
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* LLM warmup logic
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* multi-modality warmup
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* optimize code
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* refine log and fix some issue
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix warmup issue for mllama
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* pingpong optimization
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* match the latest vllm_extension ops
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* work with the latest vllm extension ops
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* remove block_scales which is not needed anymore
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* improve performance
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* prefill bypass graph
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* pingpong optimization issue fix
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
---------
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* add prometheus port
* fix doc
* add port for trtllm and llamacpp
* Fixing format after rebase.
---------
Co-authored-by: Nicolas Patry <patry.nicolas@protonmail.com>
* clean cuda/rocm code in hpu backend, enable flat_hpu
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix TP in pageattn
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* adjust block table in hpu to improve performance
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* enable all the model. not testet yet
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* use tensor cache in hpu graph to avoid replay issue
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* add moe support, fix qwen/mistral/mixtral crash
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix phimoe issue
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* gpt_bigcode could also go pageattn
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* enable dbrx remove some unused code
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* multi-modality initial PR
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* adjust warmup and enable vlm
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix incorrect output in qwen2 idefics if hpu graph is used
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* remove unused quantization code and enable awq/gptq int4
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix gptq issue
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* enable fp8
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* warmup prefill
remove model where pageattn is not used, set block table to None since it's not used
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* add warmup_decode
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* warmup decode
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* remove block_tables and prefill_cache_indices which will lead to dynamic shape
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix comment
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* missing gptj change...
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix some issue
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* remove torch.where to fix incorrect output in hpu graph model
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* match the latest vllm_extension ops
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
---------
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* feat(gaudi): release ready (docs, docker image and vlm ready)
* fix(gaudi): add default argument for the dockerfile
* fix(gaudi): remove use of latest for gaudi docker image + redid gaudi benchmarking section to include best practices
* feat(neuron): use AWS Neuron SDK 2.21.1
* feat(neuron): bump optimum-neuron version
* feat(neuron): tag latest image for local tests
* test(neuron): simplify sampling test
* Making `tool_calls` a vector.
* Update doc.
* Fixing the nix overlay with updated version.
* Add openai dependency.
* Updating the old tests.
* Trying to reduce the logs in the case of errors.
* Less spammy logs too.
* feat: add support for HF_HUB_USER_AGENT_ORIGIN to add user-agent Origin field in Hub requests.
* fix: Rust version for Neuron
* fix: PR comments, use rust-toolchain.toml
* wip(gaudi): import server and dockerfile from tgi-gaudi fork
* feat(gaudi): new gaudi backend working
* fix: fix style
* fix prehooks issues
* fix(gaudi): refactor server and implement requested changes
* feat: add neuron backend
* feat(neuron): add server standalone installation
* feat(neuron): add server and integration tests
* fix(neuron): increase ulimit when building image
The base image used to compile the rust components seems to have a low
ulimit for opened files, which leads to errors during compilation.
* test(neuron): merge integration tests and fixtures
* test: add --neuron option
* review: do not use latest tag
* review: remove ureq pinned version
* review: --privileged should be the exception
* feat: add neuron case to build ci
* fix(neuron): export models from container in test fixtures
The neuron tests require models to have been previously exported and
cached on the hub. This is done automatically by the neuron.model
fixture the first time the tests are ran for a specific version.
This fixture used to export the models using optimum-neuron directly,
but this package is not necessarily present on the system.
Instead, it is now done through the neuron TGI itself, since it
contains all the tools required to export the models.
Note that since the CI runs docker in docker (dind) it does not seem
possible to share a volume between the CI container and the container
used to export the model.
For that reason, a specific image with a modified entrypoint is built
on-the-fly when a model export is required.
* refactor: remove sagemaker entry-point
The SageMaker image is built differently anyway.
* fix(neuron): avoid using Levenshtein
* test(neuron): use smaller llama model
* feat(neuron): avoid installing CUDA in image
* test(neuron): no error anymore when requesting too many tokens
* ci: doing a precompilation step (with a different token).
* test(neuron): avoid using image sha when exporting models
We now manually evaluate the apparent hash of the neuron backend by
combining the hash of the neuron backend directory and Dockerfile.
This new hash is used to identify exported neuron models instead of the
image sha.
This has two benefits:
- it changes less frequently (only hwen the neuron backend changes),
which means less neuron models being pushed to the hub,
- it can be evaluated locally, meaning that running the tests once
locally will export the models before the CI uses them.
* test(neuron): added a small script to prune test models
---------
Co-authored-by: drbh <david.richard.holtz@gmail.com>
Co-authored-by: Nicolas Patry <patry.nicolas@protonmail.com>
