compressed-tensors is a safetensors extension for sparse, quantized
tensors. The format is more powerful than earlier AWQ/GPTQ/FP8
quantization, because
- Different quantizer configurations can be used for different targets.
- The format can specify input/output quantizers in addition to weight
quantizers.
- Configurable exclusions for quantization.
This change adds a dependency on the `compressed-tensors` package for
its configuration parsing and layer matching functionality.
The following types of quantization are supported in this PR:
- W8A16 and W4A16 INT using GPTQ-Marlin kernels.
- W8A8 and W8A16 FP using FP8-Marlin and cutlass kernels.
Support for other quantization types will be added in subsequent PRs.
* feat: support multidimensional position ids on batch to enable cuda graphs on qwen2-vl
* fix: only check model type if config exists
* fix: adjust sharding and lm head logic
* fix qwen2 failure in intel cpu
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix: return correct shape logits and add streaming test
* fix: remove unused import and refactor test
---------
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* feat: add support for qwen2 vl model
* feat: fix token padding, enable warmup and process basic request
* fix: improve get_position_ids, add lift embed_tokens
* fix: remove get_cos_sin_hack dev function
* feat: add simple test chat with meesage and text
* fix: lint test
* fix: adjust positional embeddings for multi dimensional position ids
* fix: update docs and lint unused vars
* fix: include linted file
* fix: add norm after text output
* fix: format model file
* fix: adjust for ruff lints
* fix: remove unused rotate_half
* feat: refactors and calc num features
* fix: prefer position_ids passed from vlm causal lm and reset ids on batch
* fix: adjust get_position_ids if not available and add required args to signatures
* fix: adjust resize case for qwen2_vl warmup
* fix: avoid qwen2 vl specific paths with qwen2
* Switch from fbgemm-gpu w8a8 scaled matmul to vLLM/marlin-kernels
Performance and accuracy of these kernels are on par (tested with Llama
70B and 405B). Removes a dependency and resolves some stability issues
we have been seeing.
* Update test snapshots
* Add support for FP8 KV cache scales
Since FP8 only has limited dynamic range, we can scale keys/values
before storing them into the cache (and unscale them in attention). To
avoid rescaling the cache as the absmax values change, good scales are
usually determined per layer using calibration calibration data and stored
in the checkpoint.
This change adds support for for using key-value scales and loading them
from checkpoints in the two most common formats:
- Separate per-layer `k_scale` and `v_scale` scalars.
- Per-layer `kv_scale` scalar (older format).
Currently, scales are only used with an `float8_e4m3fn` cache.
Besides adding support for key/value scales, the `fp8_quantize` function
is also extended to support quantization with a kernel vendored from
vLLM. This is slightly faster than the PyTorch implementation, but also
scales in FP32, potentially improving accuracy.
* Update FP8 KV cache test to use checkpoint with scales
* `can_scale`: check that the attention is flashinfer
Update the Mixtral GPTQ test to use a model with `desc_act=true` and
`group_size!=-1` to ensure that we are checking activation
sorting/non-full K (with tensor parallelism). The `desc_act=false` case
is already checked by the Mixtral AWQ test.
* add gptq and awq int4 support in intel platform
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* fix ci failure
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* set kv cache dtype
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* refine the code according to the review command
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
* Simplifying conditionals + reverting integration tests values.
* Unused import
* Fix redundant import.
* Revert change after rebase.
* Upgrading the tests (TP>1 fix changes to use different kernels.)
* Update server/text_generation_server/layers/gptq/__init__.py
---------
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>
Co-authored-by: Wang, Yi A <yi.a.wang@intel.com>
* feat: process token stream before returning to client
* fix: expect content in test
* fix: improve comparison via ruff lint
* fix: return event in all cases
* fix: always send event on error, avoid unwraps, refactor and improve tests
* fix: prefer no_tool over notify_error to improve reponse
* fix: adjust chat input test for no_tool
* fix: adjust test expected content
---------
Co-authored-by: System administrator <root@ip-10-90-0-186.ec2.internal>
* 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: unroll notify_error if no tool is choosen
* fix: expect simple message when no tool is selected
* fix: improve test to avoid notify_error
* fix: improve docs and indicate change in expected response
* fix: adjust linting in test file
* Working loading state.
* Preprocessing.
* Working state ? (Broke idefics1 temporarily).
* Cleaner condition.
* Fix idefics.
* Updating config, removing TODO
* Mllama
* Ugrade transformers 4.45
* Flashing mllama.
* Starting to get there.
* Working state.
* Integrations tests for mllama (cutting to 10 tokens because there seems'
to be instability after (meaning size of the batch matters.
* Updating model link.
* Earlier assert.
* Fix vlm ?
* remove log.
* Force ignore all images but last.
* Default dtype bfloat16.
* Update integration test after switch to bf16.
* Remove dead code.
* Removed dead code.
* Upgrade the flake to latest transformers/tokenizers
* Move to hf tgi-nix
* Upgrade to 0.5.0
* 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.
* Stream options.
* Fetch stuff from nix integration test for easier testing.
* Adding the assert.
* Only send the usage when asked for.
* Update the docs.
* Impure test because we need network.
* develop.
* Optional usage.
* Fixes.
* Workflow
* 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
* Adding a test for FD.
* Fixing flashdecoding (empty batch doesn't work).
* Fixing the invalid popping.
* Fixing radix with block_size > 1
* Last reference.
* Use an actual hash.
* Update hash for slice.len() == 1
* Update the locks.
* Increasing docker timeout.
* Adding prefix test.
* [WIP] tmp dump of integration load tests.
* Remove other tensor creation.
* Fixed the radix tree.
Used a slice everywhere in radix.rs to keep the cheap Arc cloning
instead of recomputing the input_ids.
* Fix parsing
* Is it really flashinfer version ?
* Remove some comments.
* Revert the max prefix hit.
* Adding numpy to diff.
* Upgraded flashinfer.
* Upgrading some stuff.
* Are we done yet ?
* Minor fixup
* Remove 1 log and put back the other.
* Add comment for why slot 0 is OK.
* Mounting on the job.
* Get me a debug branch
* Debugging CIs is fun.
* Attempt #28
* wip
* Tmate.
* Praying.
* Updating VLM causal model with updated context.
* Important line got squashed.
* Tmate again.
* Fingers crossed.
* We want only 1 run of integration tests.....
---------
Co-authored-by: Guillaume LEGENDRE <glegendre01@gmail.com>
* 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>
* All integration tests back everywhere (too many failed CI).
* Upgrade integration tests after 12.4
* Attempt to remove the specifed compute cap.
* Common arch list.
* Punica uses raw ASM which is not valid on 9.0 apparently.
* 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
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
* feat: simple mistral lora integration tests
* fix: include args in docker launcher
* fix: disable cuda graphs with lora and warn
* fix: adjust docs and precommit issues
* fix: re update docs
* Add more representative Llama GPTQ test
The Llama GPTQ test is updated to use a model with the commonly-used
quantizer config format and activation sorting. The old test is
kept around (but renamed) since it tests the format produced by
`text-generation-server quantize`.
* Add support for manually triggering a release build
* Refactor dead code.
* First working step.
* Remove a lot of duplicated code.
* More dead code.
* More cleanup.
* Fix Santacoder test.
* Fixing the simple tests.
* Fixing sharding.
* Fixes for VLM.
* Fixing santacoder (num_kv_heads hardcoded).
* Removing more dead code.
* Fixing `config.n_head`.
* Stopping earlier because of `<end_of_utterance>` in idefics2.
* Addresses comments.
* Removing the dead code.
* Fuse back mistral into FlashCausalLM.
* Finish removal.
* Fixing docs + causal_lm `batch_class`.
* Fixing docs + causal.lm.
* Add default to Gemma Causality.
* Default value for gemma/gemma2.
* Wrong default.
GPTQ-Marlin is currently the best-performing kernel for GPTQ models. So
let's use it by default if the kernels are installed, the GPU supports
it, and the kernels support the configuration.
For models generated by `text-generation-server quantize`, use
`sym=False`. This subcommand symmetric quantization since the beginning
and incorrectly reporting the model to be symmetric will use
GPTQ-Marlin (which does not support asymmetric quantization).
Before this change, the number of reserved image tokens was not the
same as the number of images. Fixes#2029.
While at it, also remove all the image token handling duplication
in `prepare_input`.
When a batch contained images if different sizes during prefill, the
server would fail (see e.g. #2056). Images were processed separately and
then concatenated. However, this can fail for images with different sizes.
Fix this by preprocessing all images in the batch together, so that the
image processor can ensure that all image tensors have compatible sizes.
Add support for GPTQ Marlin kernels
GPTQ Marlin extends the Marlin kernels to support common GPTQ
configurations:
- bits: 4 or 8
- groupsize: -1, 32, 64, or 128
- desc_act: true/false
Using the GPTQ Marlin kernels requires repacking the parameters in the
Marlin quantizer format.
The kernels were contributed by Neural Magic to VLLM. We vendor them
here for convenience.
This change adds support for Marlin-quantized models. Marlin is an
FP16xINT4 matmul kernel, which provides good speedups decoding batches
of 16-32 tokens. It supports quantized models with symmetric
quantization, groupsize -1 or 128, and 4-bit.
Tested with:
- Llama 2
- Llama 3
- Phi 3
Daniël de Kok