Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Model] MLPSpeculator speculative decoding support #4947

Merged
merged 42 commits into from
Jun 21, 2024

Conversation

JRosenkranz
Copy link
Contributor

@JRosenkranz JRosenkranz commented May 21, 2024

This is a PR to include MLPSpeculator from fms. This feature will improve inference performance significantly with minimal effort from the user by leveraging the existing speculative decoding framework with an improved speculator architecture implementation. The architecture itself is similar to medusa, with a key difference that each head conditions its outputs on the prior head. This allows it to produce better formed n-grams and increase acceptance rate. A more comprehensive description of the speculator can be found in the following technical report.

Current Speculators:

Design:

MLPSpeculator: A speculator model in modeling code which each forward pass corresponds to a single layer of the MLPSpeculator -> implement forward via single steps of the speculator.
HiddenStatesWorker: A special form of Worker which will save the hidden states -> will contain an instance of MLPSpeculator as well as the base model. Overrides the Worker execute_model method and sets the previous hidden states in the MLPSpeculator
SingleStepSpeculativeModelRunner: A model runner which will run a single step of the speculator while also only computing the inputs a single time.

This PR will leverage the work already done in spec_decode_worker. For this speculator -

proposer_worker - MultiStepWorker which is backed by an MLPSpeculator
scorer_worker - HiddenStatesWorker

This PR had a few main challenges that were addressed in the following way:

A. Implement forward in speculator such that it knows which head to be using

Because we wanted to re-use the multi_step_worker, we opted to make the MLPSpeculator model its own model in modeling code which would only return one hidden states per iteration in the multi_step_worker. To do this, we had to do the following:

1. Each time a new hidden state comes in to the HiddenStatesWorker set a counter (corresponding to the current head in the speculator) to 0
2. Each forward pass of the speculator within the multi_step_worker will increase that counter by 1 and call the proper head
3. Forward will return the hidden states and save them for the next head (previous_hidden_states)

B. Pass the hidden states to the MLPSpeculator from the scorer_worker

Our speculator requires the base model hidden states as input, which is not currently supported in vLLM. For this, we created a new HiddenStatesWorker which acts similarly to a regular worker, but instead also saves the previous hidden states within the MLPSpeculator for later use. This is done in the spec_decode_worker when score_proposals is called (in turn calling execute_model). Aside from the previous_hidden_states, this will also update any other information required for MLPSpeculator such as the context length per request id as well as the current head counter described in A.1

C. The hidden states must be pruned for next get_spec_proposals call

Because the hidden states from the base model come to the speculator unpruned, we must prune them prior to generating our new candidate tokens. Pruning will happen by getting the number of accepted tokens per request using the prior forward’s context length and the current context length (saved as instance variables in MLPSpeculator)

Using the offline example:

batch_size=1

Non-Speculative: 100%|██████████| 1/1 [00:00<00:00,  1.62it/s, Generation Speed: 86.03 toks/s]
Speculative: 100%|██████████| 1/1 [00:00<00:00,  2.52it/s, Generation Speed: 133.83 toks/s]

@cadedaniel
Copy link
Collaborator

This is awesome! I will take a look today or tomorrow.

@cadedaniel cadedaniel self-assigned this May 21, 2024
@JRosenkranz JRosenkranz marked this pull request as ready for review May 22, 2024 19:05
Copy link
Collaborator

@cadedaniel cadedaniel left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks. I am still reviewing. Initial thoughts:

  • We want to eventually support chunked prefill + spec decode + MLPSpeculator. This means that how we determine if we're in a prompt run or not will change, we should make sure the design fits it (right now I think it's missing something).
  • Do we need to re-use the multi step worker? I would support having one specialized to MLPSpeculator. The reason why is there is a lot of overhead in each step of the multi step worker (prepare inputs). If we have a custom MLPSpeculatorWorker (or similar), we can skip that for subsequent steps, and potentially even cuda graph them in the future.
  • TODO see if there's a better way to get hidden states from worker via alternative state tracking.

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
@tdoublep
Copy link
Member

tdoublep commented May 27, 2024

@cadedaniel I did a pass through and refactored things a bit to try a reduce the overhead - I now introduced a standalone MLPSpeculatorWorker so we don't need to use MultiStepWorker at all here.

Here are some performance results for llama-2-13b:

Without speculation (5 repetitions, batch size 1):

Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:04<00:00,  4.73s/it, Generation Speed: 42.32 toks/s]
0.023635914325714113
Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:04<00:00,  4.73s/it, Generation Speed: 42.30 toks/s]
0.023645806312561034
Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:04<00:00,  4.73s/it, Generation Speed: 42.29 toks/s]
0.02365394473075867
Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:04<00:00,  4.73s/it, Generation Speed: 42.27 toks/s]
0.023662294149398803
Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:04<00:00,  4.74s/it, Generation Speed: 42.21 toks/s]
0.023697214126586916

We see ~23ms per token latency and 42 toks/s throughput.

With speculation (5 repetitions, batch size 1, before refactor @ 6ba9a1e):

Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00,  2.53s/it, Generation Speed: 78.93 toks/s]
0.01267344355583191
Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00,  2.54s/it, Generation Speed: 78.90 toks/s]
0.01267833948135376
Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00,  2.54s/it, Generation Speed: 78.82 toks/s]
0.01269181728363037
Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00,  2.54s/it, Generation Speed: 78.78 toks/s]
0.012697651386260986
Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00,  2.55s/it, Generation Speed: 78.54 toks/s]
0.012735981941223145

We see ~12.7 ms per token and ~78.8 tok/sec throughput.

With speculation (5 repetitions, batch size 1, after refactor @ bf2f102):

Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00,  2.50s/it, Generation Speed: 80.15 toks/s]
0.012481423616409302
Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00,  2.50s/it, Generation Speed: 80.10 toks/s]
0.01248981237411499
Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00,  2.49s/it, Generation Speed: 80.20 toks/s]
0.012472513914108276
Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00,  2.50s/it, Generation Speed: 80.12 toks/s]
0.012485114336013793
Processed prompts: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00,  2.50s/it, Generation Speed: 80.05 toks/s]
0.01249562382698059

We see ~12.4 ms per token latency and 80.1 toks/s throughput.

  • We see a small but non-neglible improvement from refactoring via MLPSpeculatorWorker.
  • These speed-ups of nearly 2x over without speculation are pretty consistent with what we've seen from this speculator in our past experiments.

…ulator

# Conflicts:
#	vllm/model_executor/models/__init__.py
#	vllm/spec_decode/spec_decode_worker.py
@LiuXiaoxuanPKU
Copy link
Collaborator

  • TODO see if there's a better way to get hidden states from worker via alternative state tracking.

I have the same concern about this. I feel synchronizing the HiddenStatesWorker and Worker's execute model logic might introduce some overhead. Is it possible to reuse the Worker, maybe add some variables to track if we need the hidden states?

Copy link
Collaborator

@LiuXiaoxuanPKU LiuXiaoxuanPKU left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Performance numbers are really promising!
Just went a first pass, will take another pass after merging the conflict.
On the high level, why do we need to introduce the MLPSpeculatorModelRunner?

examples/offline_inference.py Outdated Show resolved Hide resolved
examples/offline_inference.py Outdated Show resolved Hide resolved
vllm/model_executor/models/mlp_speculator.py Outdated Show resolved Hide resolved
vllm/model_executor/models/mlp_speculator.py Outdated Show resolved Hide resolved
vllm/model_executor/models/mlp_speculator.py Outdated Show resolved Hide resolved
vllm/model_executor/models/mlp_speculator.py Outdated Show resolved Hide resolved
vllm/spec_decode/spec_decode_worker.py Outdated Show resolved Hide resolved
Copy link
Contributor Author

@JRosenkranz JRosenkranz left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

a few more review comments to consider

examples/offline_inference.py Outdated Show resolved Hide resolved
examples/offline_inference.py Outdated Show resolved Hide resolved
examples/offline_inference.py Outdated Show resolved Hide resolved
vllm/model_executor/models/mlp_speculator.py Outdated Show resolved Hide resolved
vllm/model_executor/models/mlp_speculator.py Outdated Show resolved Hide resolved
alpha=self.emb_weight / self.state_weight)

state = self.activation(self.ln[head_index](state)) # b k d
# todo: not yet supporting top_k_tokens_per_head
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Without this, we are assuming the set of top_k_tokens_per_head are each 1 for every head. We may want to see how this affects performance (acceptance rate) as typically we create a tree and traverse the tree to find the optimal candidate.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So, n_candidates is also not used? (it is assumed to be equal 1?)

vllm/spec_decode/spec_decode_worker.py Outdated Show resolved Hide resolved
@JRosenkranz
Copy link
Contributor Author

JRosenkranz commented Jun 3, 2024

@LiuXiaoxuanPKU

On the high level, why do we need to introduce the MLPSpeculatorModelRunner?

The MLPSpeculator takes different input than just a normal model in VLLM. It doesn't require any of the metadata required for paged attention, and also calls a different method from the MLPSpeculator. This was created to reduce the overhead of requiring all the parameters be computed for forward

We may be able to just merge this all into the MLPSpeculatorWorker though and remove the class

@cadedaniel
Copy link
Collaborator

Going to give another pass tomorrow.

@njhill
Copy link
Member

njhill commented Jun 5, 2024

Thanks @cadedaniel! I'll be pushing some updates shortly, best to wait for that before looking again.

@cadedaniel
Copy link
Collaborator

let me know when it's ready @njhill

@LiuXiaoxuanPKU
Copy link
Collaborator

Will give it a review tonight/tmr.

Copy link
Collaborator

@LiuXiaoxuanPKU LiuXiaoxuanPKU left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM, thanks!

# These are currently required for MLPSpeculator decoding
use_v2_block_manager=True,
enforce_eager=True,
)
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Impressive! It seems MLPSpeculator + w/o cudagraph is still much faster than Original + cudagraph.

@njhill njhill merged commit b12518d into vllm-project:main Jun 21, 2024
66 checks passed
@jamestwhedbee
Copy link
Contributor

FWIW, I just tested llama3-8b-acclerator out on an AMD MI100 using the openai entrypoint. At first, I was confused and disappointed looking at the throughput metrics (token/sec looked worse) until I found #5708.

Turns out chunks/second was being measured and looked worse but tokens per second were ~1.67x!

Do you all have any plans for a Llama 3 70B accelerator? I would test this on MI300X.

@cadedaniel
Copy link
Collaborator

@jamestwhedbee this is great. Do you know what batch size you used and the category of prompt?

@sahilsuneja1
Copy link
Contributor

@jamestwhedbee We have a preliminary version of the llama3-70b speculator, achieving a ~2x logical speedup (would be lesser in practice). We are working on improving it 🤞

@jamestwhedbee
Copy link
Contributor

jamestwhedbee commented Jun 21, 2024

@jamestwhedbee this is great. Do you know what batch size you used and the category of prompt?

@cadedaniel so at first I was using https://github.com/ray-project/llmperf at various batch sizes with the sonnet dataset. But I now wonder if I was getting incorrect tokens/second data from this approach (or maybe it was correct but as you will see the speculative metrics in the vLLM logs were also surprisingly low here). I don't think it is representative of the type of prompts I'd care about anyways.

Example command

python token_benchmark_ray.py --model "meta-llama/Meta-Llama-3-8B-Instruct" --mean-input-tokens 550 --stddev-input-tokens 150 --mean-output-tokens 150 --stddev-output-tokens 10 --max-num-completed-requests 8 --timeout 600 --num-concurrent-requests 1 --llm-api openai

And a representative log I saw doing this

Speculative metrics: Draft acceptance rate: 0.174, System efficiency: 0.286, Number of speculative tokens: 4, Number of accepted tokens: 130, Number of draft tokens tokens: 748, Number of emitted tokens tokens: 267.

I didn't understand what was going on yet so I also played around with the number of speculative tokens

 python -m vllm.entrypoints.openai.api_server --model meta-llama/Meta-Llama-3-8B-Instruct --speculative-model ibm-fms/llama3-8b-accelerator --num-speculative-tokens 1 --use-v2-block-manager --enforce-eager

which improved acceptance and "tokens/second" slightly

Speculative metrics: Draft acceptance rate: 0.347, System efficiency: 0.674, Number of speculative tokens: 1, Number of accepted tokens: 194, Number of draft tokens tokens: 559, Number of emitted tokens tokens: 753.

I switched to manually prompting single batch for detailed descriptions of historical events and started seeing better acceptance rates and this is when I noticed that even though "tokens/second" was lower, the actual stream was faster.

Speculative metrics: Draft acceptance rate: 0.528, System efficiency: 0.486, Number of speculative tokens: 4, Number of accepted tokens: 412, Number of draft tokens tokens: 780, Number of emitted tokens tokens: 474.

I'm very interested in understanding per-user tokens per second in higher batch sizes as well. I am worried about measurement errors though. When #5742 lands, I'd be happy to run some more robust testing.

@cadedaniel
Copy link
Collaborator

@jamestwhedbee thanks for all the details. I am suspicious of these results as well. I expect the MLPSpeculator to have better acceptance rates than 0.5 based on my experience training draft models. Could you run on a sharegpt dataset?

cc @njhill there may be a correctness issue (draft acceptance rate seems too low)

also there is a spec decode vLLM slack channel. if you want to join you can share results there.

@cadedaniel
Copy link
Collaborator

talked offline with @njhill . we are missing e2e tests for this PR. we'll follow up to see if we can quickly fix or if we have to revert.

@jamestwhedbee
Copy link
Contributor

@cadedaniel sure do you have a link handy? I can't find it.

@cadedaniel
Copy link
Collaborator

@jamestwhedbee send me an email at cade@anyscale.com

robertgshaw2-neuralmagic pushed a commit to neuralmagic/nm-vllm that referenced this pull request Jun 23, 2024
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>

Co-authored-by: Thomas Parnell <tpa@zurich.ibm.com>
Co-authored-by: Nick Hill <nickhill@us.ibm.com>
Co-authored-by: Davis Wertheimer <Davis.Wertheimer@ibm.com>
kzawora-intel added a commit to HabanaAI/vllm-fork that referenced this pull request Jul 2, 2024
* [Hardware][Intel] Optimize CPU backend and add more performance tips (vllm-project#4971)

Co-authored-by: Jianan Gu <jianan.gu@intel.com>

* [Docs] Add 4th meetup slides (vllm-project#5509)

* [Misc] Add vLLM version getter to utils (vllm-project#5098)

* [CI/Build] Simplify OpenAI server setup in tests (vllm-project#5100)

* [Doc] Update LLaVA docs (vllm-project#5437)

Co-authored-by: Roger Wang <ywang@roblox.com>

* [Kernel] Factor out epilogues from cutlass kernels (vllm-project#5391)

Co-authored-by: Michael Goin <michael@neuralmagic.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: zifeitong <zifei.tong@parasail.io>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-neuralmagic@users.noreply.github.com>

* [MISC] Remove FP8 warning (vllm-project#5472)

Co-authored-by: Philipp Moritz <pcmoritz@gmail.com>

* Seperate dev requirements into lint and test (vllm-project#5474)

* Revert "[Core] Remove unnecessary copies in flash attn backend" (vllm-project#5478)

* [misc] fix format.sh (vllm-project#5511)

* [CI/Build] Disable test_fp8.py (vllm-project#5508)

* [Kernel] Disable CUTLASS kernels for fp8 (vllm-project#5505)

* Add `cuda_device_count_stateless` (vllm-project#5473)

* [Hardware][Intel] Support CPU inference with AVX2 ISA (vllm-project#5452)

* [Misc] Fix arg names in quantizer script (vllm-project#5507)

* bump version to v0.5.0.post1 (vllm-project#5522)

* [CI/Build][Misc] Add CI that benchmarks vllm performance on those PRs with `perf-benchmarks` label (vllm-project#5073)

Co-authored-by: simon-mo <simon.mo@hey.com>

* [CI/Build] Disable LLaVA-NeXT CPU test (vllm-project#5529)

* [Kernel] Fix CUTLASS 3.x custom broadcast load epilogue (vllm-project#5516)

* [Misc] Fix arg names (vllm-project#5524)

* [ Misc ] Rs/compressed tensors cleanup (vllm-project#5432)

Co-authored-by: mgoin <michael@neuralmagic.com>
Co-authored-by: Dipika Sikka <dipikasikka1@gmail.com>

* [Kernel] Suppress mma.sp warning on CUDA 12.5 and later (vllm-project#5401)

* [mis] fix flaky test of test_cuda_device_count_stateless (vllm-project#5546)

* [Core] Remove duplicate processing in async engine (vllm-project#5525)

* [misc][distributed] fix benign error in `is_in_the_same_node` (vllm-project#5512)

* [Docs] Add ZhenFund as a Sponsor (vllm-project#5548)

* [Doc] Update documentation on Tensorizer (vllm-project#5471)

* [Bugfix] Enable loading FP8 checkpoints for gpt_bigcode models  (vllm-project#5460)

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>

* [Bugfix] Fix typo in Pallas backend (vllm-project#5558)

* [Core][Distributed] improve p2p cache generation (vllm-project#5528)

* Add ccache to amd (vllm-project#5555)

* [Core][Bugfix]: fix prefix caching for blockv2 (vllm-project#5364)

Signed-off-by: Lei Wen <wenlei03@qiyi.com>
Co-authored-by: Lei Wen <wenlei03@qiyi.com>

* [mypy] Enable type checking for test directory (vllm-project#5017)

* [CI/Build] Test both text and token IDs in batched OpenAI Completions API (vllm-project#5568)

* [misc] Do not allow to use lora with chunked prefill. (vllm-project#5538)

Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>

* add gptq_marlin test for bug report vllm-project#5088 (vllm-project#5145)

* [BugFix] Don't start a Ray cluster when not using Ray (vllm-project#5570)

* [Fix] Correct OpenAI batch response format (vllm-project#5554)

* Add basic correctness 2 GPU tests to 4 GPU pipeline (vllm-project#5518)

* [CI][BugFix] Flip is_quant_method_supported condition (vllm-project#5577)

* [build][misc] limit numpy version (vllm-project#5582)

* [Doc] add debugging tips for crash and multi-node debugging (vllm-project#5581)

* Fix w8a8 benchmark and add Llama-3-8B (vllm-project#5562)

* [Model] Rename Phi3 rope scaling type (vllm-project#5595)

* Correct alignment in the seq_len diagram. (vllm-project#5592)

Co-authored-by: Liqian Chen <liqian.chen@deeplang.ai>

* [Kernel] `compressed-tensors` marlin 24 support (vllm-project#5435)

* [Misc] use AutoTokenizer for benchmark serving when vLLM not installed (vllm-project#5588)

* [Hardware][Intel GPU] Add Intel GPU(XPU) inference backend (vllm-project#3814)

Co-authored-by: Jiang Li <jiang1.li@intel.com>
Co-authored-by: Abhilash Majumder <abhilash.majumder@intel.com>
Co-authored-by: Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>

* [CI/BUILD] Support non-AVX512 vLLM building and testing (vllm-project#5574)

* [CI] the readability of benchmarking and prepare for dashboard (vllm-project#5571)

[CI] Improve the readability of performance benchmarking results and prepare for upcoming performance dashboard (vllm-project#5571)

* [bugfix][distributed] fix 16 gpus local rank arrangement (vllm-project#5604)

* [Optimization] use a pool to reuse LogicalTokenBlock.token_ids (vllm-project#5584)

* [Bugfix] Fix KV head calculation for MPT models when using GQA (vllm-project#5142)

* [Fix] Use utf-8 encoding in entrypoints/openai/run_batch.py (vllm-project#5606)

* [Speculative Decoding 1/2 ] Add typical acceptance sampling as one of the sampling techniques in the verifier (vllm-project#5131)

* [Model] Initialize Phi-3-vision support (vllm-project#4986)

* [Kernel] Add punica dimensions for Granite 13b (vllm-project#5559)

Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>

* [misc][typo] fix typo (vllm-project#5620)

* [Misc] Fix typo (vllm-project#5618)

* [CI] Avoid naming different metrics with the same name in performance benchmark (vllm-project#5615)

* [bugfix][distributed] improve p2p capability test (vllm-project#5612)

[bugfix][distributed] do not error if two processes do not agree on p2p capability (vllm-project#5612)

* [Misc] Remove import from transformers logging (vllm-project#5625)

* [CI/Build][Misc] Update Pytest Marker for VLMs (vllm-project#5623)

* [ci] Deprecate original CI template (vllm-project#5624)

Signed-off-by: kevin <kevin@anyscale.com>

* [Misc] Add OpenTelemetry support (vllm-project#4687)

This PR adds basic support for OpenTelemetry distributed tracing.
It includes changes to enable tracing functionality and improve monitoring capabilities.

I've also added a markdown with print-screens to guide users how to use this feature. You can find it here

* [Misc] Add channel-wise quantization support for w8a8 dynamic per token activation quantization (vllm-project#5542)

* [ci] Setup Release pipeline and build release wheels with cache (vllm-project#5610)

Signed-off-by: kevin <kevin@anyscale.com>

* [Model] LoRA support added for command-r (vllm-project#5178)

* [Bugfix] Fix for inconsistent behaviour related to sampling and repetition penalties  (vllm-project#5639)

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>

* [Doc] Added cerebrium as Integration option (vllm-project#5553)

* [Bugfix] Fix CUDA version check for mma warning suppression (vllm-project#5642)

* [Bugfix] Fix w8a8 benchmarks for int8 case (vllm-project#5643)

* [Bugfix] Fix Phi-3 Long RoPE scaling implementation (vllm-project#5628)

* [Bugfix] Added test for sampling repetition penalty bug. (vllm-project#5659)

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>

* [Bugfix][CI/Build][AMD][ROCm]Fixed the cmake build bug which generate garbage on certain devices (vllm-project#5641)

* [misc][distributed] use 127.0.0.1 for single-node (vllm-project#5619)

* [Model] Add FP8 kv cache for Qwen2 (vllm-project#5656)

* [Bugfix] Fix sampling_params passed incorrectly in Phi3v example (vllm-project#5684)

* [Misc]Add param max-model-len in benchmark_latency.py (vllm-project#5629)

* [CI/Build] Add tqdm to dependencies (vllm-project#5680)

* [ci] Add A100 queue into AWS CI template (vllm-project#5648)

Signed-off-by: kevin <kevin@anyscale.com>

* [Frontend][Bugfix] Fix preemption_mode -> preemption-mode for CLI arg in arg_utils.py (vllm-project#5688)

* [ci][distributed] add tests for custom allreduce (vllm-project#5689)

* [Bugfix] AsyncLLMEngine hangs with asyncio.run (vllm-project#5654)

* [Doc] Update docker references (vllm-project#5614)

Signed-off-by: Rafael Vasquez <rafvasq21@gmail.com>

* [Misc] Add per channel support for static activation quantization; update w8a8 schemes to share base classes (vllm-project#5650)

* [ci] Limit num gpus if specified for A100 (vllm-project#5694)

Signed-off-by: kevin <kevin@anyscale.com>

* [Misc] Improve conftest (vllm-project#5681)

* [Bugfix][Doc] FIx Duplicate Explicit Target Name Errors (vllm-project#5703)

* [Kernel] Update Cutlass int8 kernel configs for SM90 (vllm-project#5514)

Co-authored-by: Varun Sundar Rabindranath <varun@neuralmagic.com>

* [Model] Port over CLIPVisionModel for VLMs (vllm-project#5591)

* [Kernel] Update Cutlass int8 kernel configs for SM80 (vllm-project#5275)

Co-authored-by: Varun Sundar Rabindranath <varun@neuralmagic.com>

* [Bugfix] Fix the CUDA version check for FP8 support in the CUTLASS kernels (vllm-project#5715)

* [Frontend] Add FlexibleArgumentParser to support both underscore and dash in names (vllm-project#5718)

* [distributed][misc] use fork by default for mp (vllm-project#5669)

* [Model] MLPSpeculator speculative decoding support (vllm-project#4947)

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>

Co-authored-by: Thomas Parnell <tpa@zurich.ibm.com>
Co-authored-by: Nick Hill <nickhill@us.ibm.com>
Co-authored-by: Davis Wertheimer <Davis.Wertheimer@ibm.com>

* [Kernel] Add punica dimension for Qwen2 LoRA (vllm-project#5441)

* [BugFix] Fix test_phi3v.py (vllm-project#5725)

* [Bugfix] Add  fully sharded layer for QKVParallelLinearWithLora (vllm-project#5665)

Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>

* [Core][Distributed] add shm broadcast (vllm-project#5399)

Co-authored-by: Cody Yu <hao.yu.cody@gmail.com>

* [Kernel][CPU] Add Quick `gelu` to CPU (vllm-project#5717)

* [Doc] Documentation on supported hardware for quantization methods (vllm-project#5745)

* [BugFix] exclude version 1.15.0 for modelscope (vllm-project#5668)

* [ci][test] fix ca test in main (vllm-project#5746)

* [LoRA] Add support for pinning lora adapters in the LRU cache (vllm-project#5603)

* [CI][Hardware][Intel GPU] add Intel GPU(XPU) ci pipeline (vllm-project#5616)

* [Model] Support Qwen-VL and Qwen-VL-Chat models with text-only inputs (vllm-project#5710)

Co-authored-by: Roger Wang <ywang@roblox.com>

* [Misc] Remove vllm-project#4789 workaround left in vllm/entrypoints/openai/run_batch.py (vllm-project#5756)

* [Bugfix] Fix pin_lora error in TPU executor (vllm-project#5760)

* [Docs][TPU] Add installation tip for TPU (vllm-project#5761)

* [core][distributed] improve shared memory broadcast (vllm-project#5754)

* [BugFix] [Kernel] Add Cutlass2x fallback kernels (vllm-project#5744)

Co-authored-by: Varun Sundar Rabindranath <varun@neuralmagic.com>

* [Distributed] Add send and recv helpers (vllm-project#5719)

* [Bugfix] Add phi3v resize for dynamic shape and fix torchvision requirement (vllm-project#5772)

* [doc][faq] add warning to download models for every nodes (vllm-project#5783)

* post-rebase api adjustments

* [Doc] Add "Suggest edit" button to doc pages (vllm-project#5789)

* [Doc] Add Phi-3-medium to list of supported models (vllm-project#5788)

* [Bugfix] Fix FlexibleArgumentParser replaces _ with - for actual args (vllm-project#5795)

* [ci] Remove aws template (vllm-project#5757)

Signed-off-by: kevin <kevin@anyscale.com>

* [Doc] Add notice about breaking changes to VLMs (vllm-project#5818)

* [Speculative Decoding] Support draft model on different tensor-parallel size than target model (vllm-project#5414)

* add pin_lora to habana components

* add WA for model loader

* fix api mismatches with ray

* tensor parallel fixes

* workers cpu alignment fix

* [Misc] Remove useless code in cpu_worker (vllm-project#5824)

* prefill/decode metadata fixes

* [Core] Add fault tolerance for `RayTokenizerGroupPool` (vllm-project#5748)

* re-enable attn metadata trimming

* worker_use_ray fix

* [doc][distributed] add both gloo and nccl tests (vllm-project#5834)

* [CI/Build] Add unit testing for FlexibleArgumentParser (vllm-project#5798)

* [Misc] Update `w4a16` `compressed-tensors` support to include `w8a16` (vllm-project#5794)

* [Hardware][TPU] Refactor TPU backend (vllm-project#5831)

* [Hardware][AMD][CI/Build][Doc] Upgrade to ROCm 6.1, Dockerfile improvements, test fixes (vllm-project#5422)

* [Hardware][TPU] Raise errors for unsupported sampling params (vllm-project#5850)

* [CI/Build] Add E2E tests for MLPSpeculator (vllm-project#5791)

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>

* [Bugfix] Fix assertion in NeuronExecutor (vllm-project#5841)

* [Core] Refactor Worker and ModelRunner to consolidate control plane communication (vllm-project#5408)

Signed-off-by: Stephanie Wang <swang@cs.berkeley.edu>
Signed-off-by: Stephanie <swang@anyscale.com>
Co-authored-by: Stephanie <swang@anyscale.com>

* [Misc][Doc] Add Example of using OpenAI Server with VLM (vllm-project#5832)

* [bugfix][distributed] fix shm broadcast when the queue size is full (vllm-project#5801)

* [Bugfix] Fix embedding to support 2D inputs (vllm-project#5829)

* [Bugfix][TPU] Fix KV cache size calculation (vllm-project#5860)

* [CI/Build] Refactor image test assets (vllm-project#5821)

* [Kernel] Adding bias epilogue support for `cutlass_scaled_mm` (vllm-project#5560)

Co-authored-by: Chih-Chieh-Yang <7364402+cyang49@users.noreply.github.com>
Co-authored-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>

* [Frontend] Add tokenize/detokenize endpoints (vllm-project#5054)

* [Hardware][TPU] Support parallel sampling & Swapping (vllm-project#5855)

* [Bugfix][TPU] Fix CPU cache allocation (vllm-project#5869)

* Support CPU inference with VSX PowerPC ISA (vllm-project#5652)

* [doc] update usage of env var to avoid conflict (vllm-project#5873)

* [Misc] Add example for LLaVA-NeXT (vllm-project#5879)

* [BugFix] Fix cuda graph for MLPSpeculator (vllm-project#5875)

Co-authored-by: Abhinav Goyal <abhinav.goyal@flipkart.com>

* [Doc] Add note about context length in Phi-3-Vision example (vllm-project#5887)

* [VLM][Bugfix] Make sure that `multi_modal_kwargs` is broadcasted properly (vllm-project#5880)

Signed-off-by: Xiaowei Jiang <xwjiang2010@gmail.com>

* [Model] Add base class for LoRA-supported models (vllm-project#5018)

* [Bugfix] Fix img_sizes Parsing in Phi3-Vision (vllm-project#5888)

* [CI/Build] [1/3] Reorganize entrypoints tests (vllm-project#5526)

* add collective crash WA

* add comment to the weird mark_step

* [Model][Bugfix] Implicit model flags and reenable Phi-3-Vision (vllm-project#5896)

* [doc][misc] add note for Kubernetes users (vllm-project#5916)

* [BugFix] Fix `MLPSpeculator` handling of `num_speculative_tokens` (vllm-project#5876)

* [BugFix] Fix `min_tokens` behaviour for multiple eos tokens (vllm-project#5849)

* [CI/Build] Fix Args for `_get_logits_warper` in Sampler Test (vllm-project#5922)

* [Model] Add Gemma 2 (vllm-project#5908)

* [core][misc] remove logical block (vllm-project#5882)

* [Kernel][ROCm][AMD] fused_moe Triton configs v2 for mi300X (vllm-project#5932)

* [Hardware][TPU] Optimize KV cache swapping (vllm-project#5878)

* [VLM][BugFix] Make sure that `multi_modal_kwargs` can broadcast properly with ring buffer. (vllm-project#5905)

Signed-off-by: Xiaowei Jiang <xwjiang2010@gmail.com>
Co-authored-by: Roger Wang <ywang@roblox.com>

* [Bugfix][Hardware][Intel CPU] Fix unpassed multi_modal_kwargs for CPU runner (vllm-project#5956)

* [Core] Registry for processing model inputs (vllm-project#5214)

Co-authored-by: ywang96 <ywang@roblox.com>

* Unmark fused_moe config json file as executable (vllm-project#5960)

* [Hardware][Intel] OpenVINO vLLM backend (vllm-project#5379)

* [Bugfix] Better error message for MLPSpeculator when `num_speculative_tokens` is set too high (vllm-project#5894)

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>

* [CI/Build] [2/3] Reorganize entrypoints tests (vllm-project#5904)

* [Distributed] Make it clear that % should not be in tensor dict keys. (vllm-project#5927)

Signed-off-by: Xiaowei Jiang <xwjiang2010@gmail.com>

* [Spec Decode] Introduce DraftModelRunner (vllm-project#5799)

* [Bugfix] Fix compute datatype for cutlass 3.x epilogues (vllm-project#5931)

* [ Misc ] Remove `fp8_shard_indexer` from Col/Row Parallel Linear (Simplify Weight Loading) (vllm-project#5928)

Co-authored-by: Robert Shaw <rshaw@neuralmagic>

* [ Bugfix ] Enabling Loading Models With Fused QKV/MLP on Disk with FP8 (vllm-project#5921)

Co-authored-by: Robert Shaw <rshaw@neuralmagic>

* Support Deepseek-V2 (vllm-project#4650)

Co-authored-by: Philipp Moritz <pcmoritz@gmail.com>

* [Bugfix] Only add `Attention.kv_scale` if kv cache quantization is enabled (vllm-project#5936)

* Unmark more files as executable (vllm-project#5962)

* [Bugfix] Fix Engine Failing After Invalid Request - AsyncEngineDeadError (vllm-project#5963)

Co-authored-by: Robert Shaw <rshaw@neuralmagic>

* [Kernel] Flashinfer for prefill & decode, with Cudagraph support for decode (vllm-project#4628)

Co-authored-by: LiuXiaoxuanPKU <llilyliupku@gmail.com>, bong-furiosa <bongwon.jang@furiosa.ai>

* [Bugfix][TPU] Fix TPU sampler output (vllm-project#5978)

* [Bugfix][TPU] Fix pad slot id (vllm-project#5977)

* [Bugfix] fix missing last itl in openai completions benchmark (vllm-project#5926)

* [Misc] Extend vLLM Metrics logging API (vllm-project#5925)

Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>

* [Kernel] Add punica dimensions for Granite 3b and 8b (vllm-project#5930)

Signed-off-by: Joe Runde <joe@joerun.de>

* [Bugfix] Fix precisions in Gemma 1 (vllm-project#5913)

* [Misc] Update Phi-3-Vision Example (vllm-project#5981)

Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>

* [Bugfix] Support `eos_token_id` from `config.json` (vllm-project#5954)

* [Core] Optimize `SequenceStatus.is_finished` by switching to IntEnum (vllm-project#5974)

* [Kernel] Raise an exception in MoE kernel if the batch size is larger then 65k (vllm-project#5939)

* [ CI/Build ] Added E2E Test For Compressed Tensors (vllm-project#5839)

Co-authored-by: Michael Goin <michael@neuralmagic.com>
Co-authored-by: Robert Shaw <rshaw@neuralmagic>

* [CI/Build] Add TP test for vision models (vllm-project#5892)

* [ CI/Build ] LM Eval Harness Based CI Testing (vllm-project#5838)

Co-authored-by: Robert Shaw <rshaw@neuralmagic>

* [Bugfix][CI/Build][Hardware][AMD] Install matching torchvision to fix AMD tests (vllm-project#5949)

* [CI/Build] Temporarily Remove Phi3-Vision from TP Test (vllm-project#5989)

* [CI/Build] Reuse code for checking output consistency (vllm-project#5988)

* [CI/Build] [3/3] Reorganize entrypoints tests (vllm-project#5966)

* [ci][distributed] fix device count call

[ci][distributed] fix some cuda init that makes it necessary to use spawn (vllm-project#5991)

* [Frontend]: Support base64 embedding (vllm-project#5935)

Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>

* [Lora] Use safetensor keys instead of adapter_config.json to find unexpected modules.  (vllm-project#5909)

Co-authored-by: sang <sangcho@anyscale.com>

* [ CI ] Temporarily Disable Large LM-Eval Tests (vllm-project#6005)

Co-authored-by: rshaw@neuralmagic.com <rshaw@neuralmagic>

* [Misc] Fix `get_min_capability` (vllm-project#5971)

* [ Misc ] Refactor w8a8 to use `process_weights_after_load` (Simplify Weight Loading) (vllm-project#5940)

Co-authored-by: Robert Shaw <rshaw@neuralmagic>

* [misc][cuda] use nvml to avoid accidentally cuda initialization (vllm-project#6007)

* [Speculative Decoding 2/2 ] Integrate typical acceptance sampler into Spec Decode Worker (vllm-project#5348)

* Revert test changes

* cleanup

* llm engine cleanup

* utils.py cleanup

* custom ops refactor

* move xops to ops

* remove vllm/hpu/attn_bias.py

* whitespace fix

* revert accidental changes in rmsnorm

* Fix hpugraph hashing

* add trim_attn_metadata comment

* fix prompt bucketing:

* [ CI ] Re-enable Large Model LM Eval (vllm-project#6031)

* [doc][misc] remove deprecated api server in doc (vllm-project#6037)

* [Misc] update benchmark backend for scalellm (vllm-project#6018)

* [doc][misc] further lower visibility of simple api server (vllm-project#6041)

Co-authored-by: Simon Mo <simon.mo@hey.com>

* [Bugfix] Use RayActorError for older versions of Ray in  RayTokenizerGroupPool (vllm-project#6039)

* [Bugfix] adding chunking mechanism to fused_moe to handle large inputs (vllm-project#6029)

* add FAQ doc under 'serving' (vllm-project#5946)

* [Bugfix][Doc] Fix Doc Formatting (vllm-project#6048)

* [Bugfix] Add explicit `end_forward` calls to flashinfer (vllm-project#6044)

* [BugFix] Ensure worker model loop is always stopped at the right time (vllm-project#5987)

* [Frontend] Relax api url assertion for openai benchmarking (vllm-project#6046)

* [Model] Changes to MLPSpeculator to support tie_weights and input_scale (vllm-project#5965)

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
Co-authored-by: Joshua Rosenkranz <jmrosenk@us.ibm.com>

* [Core] Optimize block_manager_v2 vs block_manager_v1 (to make V2 default)  (vllm-project#5602)

* [Frontend] Add template related params to request (vllm-project#5709)

* [VLM] Remove `image_input_type` from VLM config (vllm-project#5852)

Signed-off-by: Xiaowei Jiang <xwjiang2010@gmail.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: Roger Wang <ywang@roblox.com>

* [Doc] Reinstate doc dependencies (vllm-project#6061)

* guard model loader wa for hpu

---------

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
Signed-off-by: Lei Wen <wenlei03@qiyi.com>
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
Signed-off-by: kevin <kevin@anyscale.com>
Signed-off-by: Rafael Vasquez <rafvasq21@gmail.com>
Signed-off-by: Stephanie Wang <swang@cs.berkeley.edu>
Signed-off-by: Stephanie <swang@anyscale.com>
Signed-off-by: Xiaowei Jiang <xwjiang2010@gmail.com>
Signed-off-by: Joe Runde <joe@joerun.de>
Co-authored-by: Li, Jiang <jiang1.li@intel.com>
Co-authored-by: Jianan Gu <jianan.gu@intel.com>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Co-authored-by: Roger Wang <ywang@roblox.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Michael Goin <michael@neuralmagic.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
Co-authored-by: zifeitong <zifei.tong@parasail.io>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-neuralmagic@users.noreply.github.com>
Co-authored-by: Cody Yu <hao.yu.cody@gmail.com>
Co-authored-by: Philipp Moritz <pcmoritz@gmail.com>
Co-authored-by: Antoni Baum <antoni.baum@protonmail.com>
Co-authored-by: Jie Fu (傅杰) <jiefu@tencent.com>
Co-authored-by: Allen.Dou <allen.dou@hotmail.com>
Co-authored-by: Simon Mo <simon.mo@hey.com>
Co-authored-by: Kuntai Du <kuntai@uchicago.edu>
Co-authored-by: Dipika Sikka <dipikasikka1@gmail.com>
Co-authored-by: Sanger Steel <sangersteel@gmail.com>
Co-authored-by: Thomas Parnell <tpa@zurich.ibm.com>
Co-authored-by: leiwen83 <leiwen83@users.noreply.github.com>
Co-authored-by: Lei Wen <wenlei03@qiyi.com>
Co-authored-by: SangBin Cho <rkooo567@gmail.com>
Co-authored-by: Alexander Matveev <59768536+alexm-neuralmagic@users.noreply.github.com>
Co-authored-by: Nick Hill <nickhill@us.ibm.com>
Co-authored-by: Amit Garg <gargamit@microsoft.com>
Co-authored-by: Charles Riggins <liqianchen123@foxmail.com>
Co-authored-by: Liqian Chen <liqian.chen@deeplang.ai>
Co-authored-by: zhyncs <me@zhyncs.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Co-authored-by: Abhilash Majumder <abhilash.majumder@intel.com>
Co-authored-by: Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
Co-authored-by: Bruce Fontaine <bruce@2.7182.net>
Co-authored-by: zifeitong <zifeitong@gmail.com>
Co-authored-by: sroy745 <142070531+sroy745@users.noreply.github.com>
Co-authored-by: Isotr0py <2037008807@qq.com>
Co-authored-by: Joe Runde <joe@joerun.de>
Co-authored-by: Chang Su <chang.s.su@oracle.com>
Co-authored-by: Roger Wang <136131678+ywang96@users.noreply.github.com>
Co-authored-by: Kevin H. Luu <kevin@anyscale.com>
Co-authored-by: Ronen Schaffer <ronen.schaffer@ibm.com>
Co-authored-by: sergey-tinkoff <167607910+sergey-tinkoff@users.noreply.github.com>
Co-authored-by: milo157 <43028253+milo157@users.noreply.github.com>
Co-authored-by: Shukant Pal <SukantK2002@outlook.com>
Co-authored-by: Hongxia Yang <62075498+hongxiayang@users.noreply.github.com>
Co-authored-by: DearPlanet <junsong.zhang2021.work@outlook.com>
Co-authored-by: Rafael Vasquez <rafvasq21@gmail.com>
Co-authored-by: Varun Sundar Rabindranath <varunsundar08@gmail.com>
Co-authored-by: Varun Sundar Rabindranath <varun@neuralmagic.com>
Co-authored-by: Joshua Rosenkranz <joshua.rosenkranz@gmail.com>
Co-authored-by: Davis Wertheimer <Davis.Wertheimer@ibm.com>
Co-authored-by: Jinzhen Lin <linjinzhen@hotmail.com>
Co-authored-by: Jee Li <pandaleefree@163.com>
Co-authored-by: rohithkrn <rohith.nallamaddi@gmail.com>
Co-authored-by: Murali Andoorveedu <37849411+andoorve@users.noreply.github.com>
Co-authored-by: Woo-Yeon Lee <wooyeonlee0@gmail.com>
Co-authored-by: Matt Wong <156021403+mawong-amd@users.noreply.github.com>
Co-authored-by: aws-patlange <90803007+aws-patlange@users.noreply.github.com>
Co-authored-by: Stephanie Wang <swang@cs.berkeley.edu>
Co-authored-by: Stephanie <swang@anyscale.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Chih-Chieh-Yang <7364402+cyang49@users.noreply.github.com>
Co-authored-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>
Co-authored-by: sasha0552 <admin@sasha0552.org>
Co-authored-by: Chip Kerchner <49959681+ChipKerchner@users.noreply.github.com>
Co-authored-by: Abhinav Goyal <abhinav.goyal@flipkart.com>
Co-authored-by: xwjiang2010 <87673679+xwjiang2010@users.noreply.github.com>
Co-authored-by: Divakar Verma <137818590+divakar-amd@users.noreply.github.com>
Co-authored-by: Ilya Lavrenov <ilya.lavrenov@intel.com>
Co-authored-by: Robert Shaw <rshaw@neuralmagic>
Co-authored-by: wangding zeng <155410488+zwd003@users.noreply.github.com>
Co-authored-by: Lily Liu <lilyliupku@gmail.com>
Co-authored-by: LiuXiaoxuanPKU <llilyliupku@gmail.com>, bong-furiosa <bongwon.jang@furiosa.ai>
Co-authored-by: mcalman <68564154+mcalman@users.noreply.github.com>
Co-authored-by: William Lin <SolitaryThinker@users.noreply.github.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: llmpros <10524065+llmpros@users.noreply.github.com>
Co-authored-by: sang <sangcho@anyscale.com>
Co-authored-by: Avshalom Manevich <12231371+avshalomman@users.noreply.github.com>
Co-authored-by: James Whedbee <jamesw@telnyx.com>
Co-authored-by: Joshua Rosenkranz <jmrosenk@us.ibm.com>
Co-authored-by: danieljannai21 <100521221+danieljannai21@users.noreply.github.com>
xjpang pushed a commit to xjpang/vllm that referenced this pull request Jul 8, 2024
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>

Co-authored-by: Thomas Parnell <tpa@zurich.ibm.com>
Co-authored-by: Nick Hill <nickhill@us.ibm.com>
Co-authored-by: Davis Wertheimer <Davis.Wertheimer@ibm.com>
@WangErXiao
Copy link

How can I get the accelerator model for a existed model?

@haichuan1221
Copy link
Contributor

@jamestwhedbee We have a preliminary version of the llama3-70b speculator, achieving a ~2x logical speedup (would be lesser in practice). We are working on improving it 🤞

Can we test the llama 3 70b speculator now?

@Alvant
Copy link
Contributor

Alvant commented Jul 9, 2024

How can I get the accelerator model for a existed model?

According to the paper, here is the source code for accelerator training. However, as far as I understood (from the training params docs here and code here), only Llama models are supported for acceleration...

@JRosenkranz
Copy link
Contributor Author

JRosenkranz commented Jul 9, 2024

How can I get the accelerator model for a existed model?

According to the paper, here is the source code for accelerator training. However, as far as I understood (from the training params docs here and code here), only Llama models are supported for acceleration...

There is no limitation on model architecture. We have an experimental branch where we have done training with GPTBigCode as well as Mixtral here. Feel free to take a look, however it is still experimental.

For the above speculators, granite-20b-code-instruct-accelerator is a speculator for a base model with GPTBigCode architecture.

@JRosenkranz
Copy link
Contributor Author

JRosenkranz commented Jul 9, 2024

@jamestwhedbee We have a preliminary version of the llama3-70b speculator, achieving a ~2x logical speedup (would be lesser in practice). We are working on improving it 🤞

Can we test the llama 3 70b speculator now?

@haichuan1221 We are working on releasing it. Will keep you updated as to when it is available. We still need end-to-end verifications, but results look promising!

@haichuan1221
Copy link
Contributor

@jamestwhedbee We have a preliminary version of the llama3-70b speculator, achieving a ~2x logical speedup (would be lesser in practice). We are working on improving it 🤞

Can we test the llama 3 70b speculator now?

@haichuan1221 We are working on releasing it. Will keep you updated as to when it is available. We still need end-to-end verifications, but results look promising!

I am looking forward for it!

xjpang pushed a commit to xjpang/vllm that referenced this pull request Jul 24, 2024
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>

Co-authored-by: Thomas Parnell <tpa@zurich.ibm.com>
Co-authored-by: Nick Hill <nickhill@us.ibm.com>
Co-authored-by: Davis Wertheimer <Davis.Wertheimer@ibm.com>
@sahilsuneja1
Copy link
Contributor

@haichuan1221: you can find the speculator for llama3-70b here: https://huggingface.co/ibm-fms/llama3-70b-accelerator

@hustxiayang
Copy link

@haichuan1221: you can find the speculator for llama3-70b here: https://huggingface.co/ibm-fms/llama3-70b-accelerator

hi, it seems that "model.safetensors " is not complete and just one shard.

@sahilsuneja1
Copy link
Contributor

sahilsuneja1 commented Aug 2, 2024

hi, it seems that "model.safetensors " is not complete and just one shard.

@hustxiayang, that's strange. I tested the model.safetensors version using this, and it worked ok.
Does, pytorch_model.bin ckpt version work for you?

@njhill I tested using FMS setup, maybe vLLM setup doesn't like auto-generated model.safetensors? Could you please confirm when you have time?

@njhill
Copy link
Member

njhill commented Aug 21, 2024

@hustxiayang how long are the prompts you’re using? There was an issue where the spec decoding gets disabled when the num of tokens is > 2k. This is now fixed and will be in the upcoming 0.5.5 release.

Temirulan pushed a commit to Temirulan/vllm-whisper that referenced this pull request Sep 6, 2024
Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>

Co-authored-by: Thomas Parnell <tpa@zurich.ibm.com>
Co-authored-by: Nick Hill <nickhill@us.ibm.com>
Co-authored-by: Davis Wertheimer <Davis.Wertheimer@ibm.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.