GitLab

Update kv_cache.py

* Add basic FP8 KV cache support

This change adds rudimentary FP8 KV cache support. The support is
enabled by passing `--kv-cache-dtype fp8_e5m2` to the launcher. Doing so
uses this type for the KV cache. However support is still limited:

* Only the `fp8_e5m2` type is supported.
* The KV cache layout is the same as `float16`/`bfloat16` (HND).
* The FP8 KV cache is only supported for FlashInfer.
* Loading of scales is not yet supported.

* Fix Cargo.toml

This change uses the updated Marlin MoE kernel from vLLM to support
MoE with activation sorting and groups.

* feat: support phi3.5 moe model loading

* fix: prefer llama base model and improve rotary logic

* feat: return reasonable generation and add integration test

* fix: run lint and update docs

* fix: rerun lint for openapi docs

* fix: prefer do_sample false unless temp is set by user, and update chat tests

* fix: small typo adjustments

* fix: consolidate long rope paths

* fix: revert greedy by default and test changes

* Vendor configuration so that we don't have to `trust_remote_code`

* Use SparseMoELayer

* Add support for dense MoE

* Some type annotations

* Add the usual model tests

* Ruff.

---------
Co-authored-by: Daniël de Kok <me@danieldk.eu>
Co-authored-by: Nicolas Patry <patry.nicolas@protonmail.com>

This change add support for MoE models that use GPTQ quantization.
Currently only models with the following properties are supported:

- No `desc_act` with tensor parallelism, unless `group_size=-1`.
- No asymmetric quantization.
- No AWQ.

* style

* update torch

* ix issues

* fix clone

* revert mkl

* added custom PA

* style

* fix style

* style

* hide env vart

* fix mixtral model

* add skinny kernel and merge fixes

* fixed style

* fix issue for sliding window models

* addressed review comments

* fix import

* improved error messag

* updated default value

* remove import

* fix imports after rebase

* float16 dep

* improve dockerfile

* cleaned dockerfile

* Improve support for GPUs with capability < 8

- For models that cannot use flashinfer, use flash-attn v1 + paged
  attention for models with a compute capability older than 8.
- Disable prefix caching when using paged attention.
- When using flash-attn v1, pass the key/value, rather than the
  cache, since v1 cannot use block tables.

* nix: add flash-attn-v1 to the server environment

* Move disabling prefix caching into the block of exceptions

* Capability as `usize`s

* More tensor cores.

* Fixing the logic.

* Gemma is modified by this.

This replaces the custom layers in both models.

* Add support for scalar FP8 weight scales

* Support LLM compressor FP8 checkpoints on H100

On H100, we use fbgemm-gpu, which requires bfloat16 as the input dtype.
However, we wouldn't pick up fp8 quantization for models quantized with
LLM compressor. This change adds enough parsing to detect if models have
FP8-quantized weights.

* Remove stray debug print

* Move to moe-kernels package and switch to common MoE layer

This change introduces the new `moe-kernels` package:

- Add `moe-kernels` as a dependency.
- Introduce a `SparseMoELayer` module that can be used by MoE
  models.
- Port over Mixtral and Deepseek.

* Make `cargo check` pass

* Update runner

enable intel ipex cpu and xpu in python3.11
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

fix regression caused by attention api change. ipex.varlen_attention does not support paged-cache
format kv input now.
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* Tied embeddings in MLP speculator.

* Fixing the scale_weight when users decide to not use the speculation as
much as defined in the config.

* Adding scaling support + optimize some ops.

* Making prefix/flashinfer the default and testing the full release tests.

* Include flashinfer in the docker.

* Using prebuilt.

* Allowing window_left_size (dummy version).

* Disabling flashinfer/prefix caching on odd head_dim

* Disable prefix caching for lora.

* More specific codes.

* Update lock

* Updating integration tests with new values with FI/FD.

Remove paged as a default too, and using FD everywhere.

* Update cargo lock ?

* Upgrade to 1.80 because of bitstream...

* Everywhere 1.80

* Forgot last default place.

* Apply suggestions from code review
Co-authored-by: drbh <david.richard.holtz@gmail.com>

* Updated flake lock

* Tmp

* Upgrade resolution system for less errors in resolution.

* Remove lambda for cleaner function.

* Handling debugger.

* OVerride the env in server tests.

* Is this enough to make it work ?

* This seems to be working.

* Downgrade some logs.

* Fixing the default for vlm.

* Don't enable prefix caching on VLM just yet.

* Change `add_special_tokens` in order to have the correct tokens for chat
input and not (since it's super important with the prefixing now)

* Fixing prefix caching for flashdecoding.

* Update all models.

* Fixed flashinfer version.

* add_special_tokens is internal only

* Fixing seqlen with the new vlms.

* Fixing the issue with `add_special_tokens` not being passed around.

* Fixing the test.

* Removing encoder_decoder (seq2seq).

* Update the chat test.

* Fixing the batching tokenization in flash causal lm.

* Truncating left for radix purposes.

* Oops this doesn't belong here.

* Put back default pure shell.

* Update server tests

- Default to throughput test in k6
- Use TGI_WIGGLE_ROOM to adjust wiggle room

* Only n_heads / process_group.size() are necessary.

* Revert the integrationt tests change (seem linked to head_size
modification).

* Adding error message when assert is violated.

* Fixing the free algorithm to handle times where the common prefix is
smaller.

* Apply suggestions from code review
Co-authored-by: OlivierDehaene <olivier@huggingface.co>

* Update server/text_generation_server/layers/attention/common.py
Co-authored-by: OlivierDehaene <olivier@huggingface.co>

* Fix disabling prefix caching - Fix windowing checks.

* Revert the Cohere tokenizer change (for now using a revision instead).

* Fmt.

---------
Co-authored-by: drbh <david.richard.holtz@gmail.com>
Co-authored-by: OlivierDehaene <olivier@huggingface.co>

* Prefix caching WIP

* Fixing prefix attention.

* Fixing flashinfer import.

* Fixing black.

* Fixing medusa (still wrong outputs, but functional).

* Just medusa values now.

* Fixing medusa without prefix caching.

* Fixing prefix caching.

* Medusa requires reshaping.

* Removing the logs.

* Remove router.nix

* Fixup:

- Remove logs
- Disable VLMs (they do not work)
- Disable prefix caching when user wants prefill logprobs.

* Update flake.lock

---------
Co-authored-by: Daniël de Kok <me@danieldk.eu>

* Upgrading exl2.

* Fixing the other pathways.

* Fix idefics.

fix: adds causal to attention params to check when using flash attn v1

* Using an enum for flash backens (paged/flashdecoding/flashinfer)

* Early exit on server too.

* Clippy.

* Fix clippy and fmt.

This change adds support for FlashInfer. FlashInfer can be enabled using
`FLASH_INFER=1` and is currently only implemented in `FlashCausalLM`.
Since this functionality is currently only for testing, FlashInfer is
not installed anywhere yet.

The FlashInfer API is quite different from FlashAttention/vLLM in that
it requires more global bookkeeping:

* A wrapper class needs to be contstructed (which we just call *state*).
  Since this is fairly expensive (due to pinned host memory allocation),
  we only do this once in a FlashCausalLM instance or for each CUDA
  Graph size.
* Each model forward call needs to be wrapped in `begin_forward` and
  `end_forward`. This sets up data structures that can be reused for all
  calls to attention for that forward call.

When calling attention, we need access to the state object. To avoid
passing an argument down the call chain (which would require changes to
all models), we use a context variable.

Each model forward call is wrapped using a context manager that does all
the bookkeeping for such a call:

* Set the context variable to the forward call's state.
* Call `begin_forward` on the state.
* Yield.
* Call `end_forward` on the state.
* Reset the context variable.

We cannot use a single shared global variable for this, since e.g. CUDA
Graphs of different sizes each have their own state.

* hotfix: fix xpu crash brought by code refine. torch.xpu rely on import ipex
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* reable gemma2 in xpu
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* fix in regression in ipex flashattention
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

---------
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
Co-authored-by: Wang, Yi A <yi.a.wang@intel.com>

* fix: attempt forward on flash attn2 to check hardware support

* fix: warn window_size_left when using flash attn 1

* fix: prefer version check over test op and avoid window_size_left if not flash attn2

* fix: improve condtional and error message

* fix: update sliding window conditional

* fix: simplify changes and revert model changes

* fix: avoid changing conditional

* fix: typo tweak

- Always return the hidden states.
- Create the output tensor inside the `attention` and `paged_attention`
  functions.

This removes the difference between how the output is handled between
attention (output parameter) and paged attention (return value). This
also removes the assumption that the attention implementation can
write to an output tensor (in preparation of FlashInfer).

The `GPTWeightLoader` was structured like this in pseudocode:

if marlin:
  Set up tensors in a way that GPTQ-Marlin expects
else:
  Set up tensors in a way that ExLlama/GPTQ/AWQ expect

However, the GPT-Marlin implementation details should really be in the
`marlin` module. So move the former part out to a separate
`GPTQMarlinWeightsLoader`.

- Create `quantization_config` option in the model config.
- Don't store the quantizer config in tensors anymore.

* feat: add ruff and resolve issue

* fix: update client exports and adjust after rebase

* fix: adjust syntax to avoid circular import

* fix: adjust client ruff settings

* fix: lint and refactor import check and avoid model enum as global names

* fix: improve fbgemm_gpu check and lints

* fix: update lints

* fix: prefer comparing model enum over str

* fix: adjust lints and ignore specific rules

* fix: avoid unneeded quantize check

* Fix GPTQ autotune data type to be compatible with Torch 2.4.0

* Update poetry lock file

* Fix small PaliGemma logprob differences after the torch update

* fix: refactor adapter weight loading and mapping

* feat: enable lora load from directory

* fix: adjust launcher for local lora adapters

* feat: improve weight loading and add tests

* fix: improve logging and rebase syntax issue

* fix: impove adapter merge comments and remove unused conditional

* fix: improve get_model_with_lora_adapters naming

* fix: comment typo

The marlin.py file was getting large, split it up.

* Add support for Llama 3 rotary embeddings

* Update transformers to 4.43

* Add support for repacking AWQ weights for GPTQ-Marlin

So far we couldn't support AWQ because virtually all AWQ models use
symmetric quantization, which GPTQ-Marlin did not suppors. GPTQ-Marlin
has recently added support AWQ repacking and AWQ asymmetric quantization
(zero_point=True).

This change updates all GPTQ-Marlin kernels from upstream and wires up
AWQ support. For now enabling AWQ using Marlin requires running TGI with
`--quantize gptq`.

* Enable Marlin for supported AWQ configurations by default

This makes the AWQ -> GPTQ repack test redundant, since we are now
testing this with the regular AWQ test.

* fix(l4): fix fp8 logic on l4

* also quant weights with single scale

* use marlin even on 89

* Softcapping for gemma2.

* Less clutter.

* No access to transformers config, only config_dict here.

* 0.0 is the null value in the C++ API.

* fix(server): fix fp8 weight loading

* fixed scales loading

* update snap

* revert default dtype

* feat(fp8): add support for fbgemm

* allow loading fp8 weights directly

* update outlines

* fix makefile

* build fbgemm

* avoid circular import and fix dockerfile

* add default dtype

* refactored weights loader

* fix auto conversion

* fix quantization config parsing

* force new nccl on install

* missing get_weights implementation

* increase timeout