* 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
* 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.
* 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
* 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
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>
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`.
* 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
* 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.
Deepseek V2 is a MoE model from Deepseek. Relevant variations
compared to other models:
- Grouped top-K in expert selection.
- mscale in yarn is calculated using the `mscale` and `mscale_all_dim`
configuration options.
- `mscale_all_dim` is also used in scaling attention softmax.
- Permuting of the query/key representations before applying rotary
embeddings.
- Some projections cannot be sharded (`q_a_proj`, `kv_a_proj_with_mqa`).
So, we need weight loads that supports quantized weights. To this
end `{Weights,WeightLoader}.get_weight` was added.
- The query/key head dimensionality differs from that of the value,
so we need to pad during attention.
- Heads with size 192, needs an extension to our paged attention
fork and we need to ensure that the KV cache is allocated with the
correct size.
- Shared experts.
* Improve the handling of quantized weights
Handling of quantized weights was split between two mechanisms:
- For quantized checkpoints, we used the new weight loader
infrastructure.
- For quantization while loading (EETQ, FP8, bitsandbytes) we
instead relied on conditional in `get_linear`.
Weight loaders support context managers to selectively load
particular layers with different weight loaders, which is useful
for models like Idefics2 AWQ, which uses a quantized text model,
but unquantized vision and connector models. However, the context
manager would be overrided by `get_linear`, which string-checks
`quantizer`. Also, the context manager would not work with
EETQ, FP8, and bitsandbytes.
This change migrates all quantizers to the weight loader infrastructure.
This has several benefits:
- We can use context managers with all quantizers.
- All the implementation details move down to the quantizer layers,
`get_linear` does not need to know how to handle quantizer linear
layers.
- All quantizer weights are strongly typed, we don't pass around
raw tensors.
- We don't have to pass around the `quantizer` string everywhere.
* Exclude non-MLP layers when using FP8 quantization with Llama
Packing of asymmetric quantization is broken, all (q)zeros values
of `0` get reset to `1`, resulting in a loss of accuracy. So instead
use symmetric quantization. To be able to distinguish models with
symmetric and asymmetric quantization, a new config tensor `gptq_sym` is
added. If this tensor is not present, we assume `sym=False`.
Use FP8 GPTQ-Marlin kernels to enable FP8 support on CUDA GPUs
with compute capability >=8.0 and <8.9.
Co-authored-by: Florian Zimmermeister <flozi00.fz@gmail.com>
Quantized weights were loaded in the `Weights` class, but this was
getting quite unwieldy, where every higher level method to load weights
was a long conditional to cover all the different quantizers.
This change moves loading of quantized weights out of the `Weights`
class. This is done by defining a simple `WeightsLoader` interface
that is implemented by `Exl2WeightsLoader`, `GPTQWeightsLoader`,
and `MarlinWeightsLoader`. These implementations are in the quantizers'
respective modules. The `Weights` class provides the low-level load
operations (such as loading tensors or sharded tensors), but delegates
loads that need quantizer-specific weight processing to a loader. The
loaders still use the low-level functionality provided by `Weights`.
I initially tried making a hierarchy where a class like `GPTQWeights`
would inherit from `Weights`. But it is not very flexible (e.g. does
not work well with the new weight storage mock used in tests) and
the implicit indirections made the code harder to follow.
Daniël de Kok