This PR fixes problems related to #569:
- block initialization
- throughput calculation and cache usage
- mixtral in tests
Beam search is removed for Mixtral and Llama for now. Those models use DynamicCache, which requires special function to change: (see https://github.com/huggingface/transformers/blob/main/src/transformers/cache_utils.py#L161)
---------
Co-authored-by: Max Ryabinin <mryabinin0@gmail.com>
This PR attempts to optimize the inference of Falcon models in the single-token setup by reducing the majority of Python overhead and making several assumptions about the setup. Specifically,
* Layer normalization, QKV projection (with splitting) and rotary embeddings are executed through CUDA graphs, which reduces most overhead related to small kernel launche
* If no sin/cos tensors are cached by the rotary embedding layer, we cache them for 8192 tokens (INFERENCE_MAX_LENGTH) during the first forward pass. In general, it should be beneficial to always run a max-length sequence before starting a block, but this is a question for another PR
The PR also adds a small test to ensure that the results (without quantization) of the block before and after quantization indeed match.
Lastly, the pull request makes the backward pass work (as discussed in https://github.com/bigscience-workshop/petals/pull/499) by making cached sin/cos for RotaryEmbedding into buffers and disabling the inference mode during their creation.
This PR makes both clients and servers work on macOS. Specifically, it:
- Follows https://github.com/learning-at-home/hivemind/pull/586 to run a macOS-compatible `p2pd` binary (both x86-64 and ARM64 are supported)
- Fixes forking issues and tests on macOS, Python 3.10+
- Introduces basic support for serving model blocks on Apple M1/M2 GPUs (torch.mps)
- Increases max number of open files by default (it's not enough on Linux and is really small on macOS)
- rpc_inference: server will now accept allocation timeout from user, defaults to no timeout
- bugfix: inference timeout is now measured from the moment the request is received
- previously, you would have to wait for your timeout plus the time it takes to sort through the queue (other users' timeout)
- now, you get AllocationFailed if you had to wait for over (timeout) seconds - regardless of other users
- a request for inference with no timeout will now fail instantly if there is not enough memory available
- dtype number of bytes is now correctly determined for int, bool & other types
---------
Co-authored-by: Your Name <you@example.com>
Co-authored-by: Alexander Borzunov <borzunov.alexander@gmail.com>
Co-authored-by: Aleksandr Borzunov <hxrussia@gmail.com>
This PR drops custom generation codes and introduces compatibility with `transformers.GenerationMixin` instead. This includes support for more sampling options (`top_p`, `top_k`, `repetition_penalty` requested in #460) and beam search - all that is now identical to running model with transformers locally.
Most features (excluding beam search and other rarely used stuff) are also compatible with resuming existing sessions.
### Breaking changes
If `.generate()` or forward passes are being run inside an `.inference_session()` context, they now use the opened session by default. So, these snippets are now equivalent:
```python
# Using default session
with model.inference_session(max_length=100):
output_ids = model.generate(input_ids, max_new_tokens=3)
# Explicitly specifying a session
with model.inference_session(max_length=100) as sess:
output_ids = model.generate(input_ids, max_new_tokens=3, session=sess)
```
Earlier, the 1st snippet was creating a new session, which is not what most people expected (= such code was most likely to introduce a bug, which is now fixed).
Right now, long inference requests may occupy Runtime for a few seconds without giving it away to process short (most latency-sensitive requests). This PR fixes it by disallowing the merged pool for long requests and prioritizing the short ones.
This PR extends CI to:
1. Test Llama code using [TinyLlama-v0](https://huggingface.co/Maykeye/TinyLLama-v0).
2. Test rebalancing (sets up a situation where the 1st server needs to change its original position).
3. Check if benchmark scripts run (in case someone breaks its code). Note that the benchmark results are meaningless here (since they're measured on a tiny swarm of CPU servers, with low `--n_steps`).
4. Test `petals.cli.run_dht`.
5. Increase swap space and watch free RAM (a common issue is that actions are cancelled without explanation if there's not enough RAM - so it's a useful reminder + debug tool).
6. Fix flapping tests for bloom-560m by increasing tolerance.
Other minor changes: fix `--help` messages to show defaults, fix docs, tune rebalancing constants.
This PR is designed to avoid OOMs when processing long sequences that happen due to the huge attention logits matrices.
Co-authored-by: Alexander Borzunov <borzunov.alexander@gmail.com>
* Hide GeneratorExit in _iterate_inference_steps()
* Update README.md about `--public_name`
* Use .from_pretrained(..., use_auth_token=token) instead of token=token
until it's fully supported across HF libs
* Use default network speed 25 Mbit/s
* Apply relay penalty in max-throughput routing
* Replace RPS with "tokens/sec per block" in logs
* Increase default expiration
Inference RPS may be very different from forward RPS. E.g., currently bnb uses a completely different algorithm for NF4 inference. We report detailed RPS info that can be then used for shortest-path routing for inference.
Implement an option to deploy PEFT adapters to a server. Clients can set active_adapter=... to use these adapters.
---------
Co-authored-by: Aleksandr Borzunov <borzunov.alexander@gmail.com>
Co-authored-by: justheuristic <justheuristic@gmail.com>
This PR:
1. **Abolishes the model conversion procedure.** Now, models are downloaded directly from original repositories like https://huggingface.co/bigscience/bloom. Servers download only shards with blocks to be hosted, and clients download only shards with input/output embeddings and layernorms.
- BLOOM is loaded from `bigscience/bloom`, but we use the DHT prefix `bigscience/bloom-petals` for backward compatibility. Same with smaller BLOOMs and BLOOMZ.
- LLaMA can be loaded from any repo like `username/llama-65b-hf`, but we use the DHT prefix `llama-65b-hf` (without the username) to accomodate blocks from different repos (there're a few of them with minor differences, such as `Llama` vs. `LLaMA` in the class name).
2. **Refactors the client to generalize it for multiple models.** Now, we have `petals.models` packages that contain model-specific code (e.g. `petals.models.bloom`, `petals.models.llama`). General code (e.g. CPU-efficient LM head, p-tuning) is kept in `petals.client`.
3. **Introduces** `WrappedLlamaBlock`, `DistributedLlamaConfig`, `DistributedLlamaForCausalLM`, `DistributedLlamaForSequenceClassification`, and `DistributedLlamaModel` compatible with Petals functionality (p-tuning, adapters, etc.).
4. **Introduces** `AutoDistributedConfig` that automatically chooses the correct config class (`DistributedLlamaConfig` or `DistributedBloomConfig`). The refactored configs contain all model-specific info for both clients and servers.
Upgrade instructions:
- Remove disk caches for blocks in old (converted) format to save disk space. That is, remove `~/.cache/petals/model--bigscience--bloom-petals` and `~/.cache/petals/model--bigscience--bloomz-petals` directories (if present).
We need to sample the next server using its throughput as the weight to actually achieve max throughput for fine-tuning.
As an example, imagine a situation where we have 3 servers with throughputs [1000, 500, 1] hosting the same blocks, then compare the uniform and weighted sampling strategies.
This PR:
1. **Extracts `SequenceManagerConfig` and `SequenceManagerState` subclasses.**
The config is provided by caller and never changed from inside `RemoteSequenceManager`. The state is a part of the `RemoteSequenceManager`'s state shared between the main manager and its slices. We fix some slicing bugs along the way.
2. **Removes `dht_prefix` and `p2p` arguments, makes `dht` argument optional.**
`dht_prefix` can always be overridden using `config.dht_prefix`. `p2p` actually needed only under the hood of `RemoteSequenceManager`, so it can extract it by itself without exposing this low-level class to callers. If strictly necessary, a caller can provide `p2p` as a part of `SequenceManagerState`. `dht` is also needed only by `RemoteSequenceManager`, so we can make it optional in the parent classes and create it automatically when it's not provided.
3. **Simplifies retry logic.**
Previously, we could have "nested" retry loops: one in `._update()`, another in inference/forward/backward steps. The loop in `._update()` could introduce issues to concurrent inference/forward/backward calls, since it blocks the entire class if its delay period becomes too high. Now this logic is simplified: `._update()` performs only one attempt to fetch the DHT info, any retries are triggered by the inference/forward/backward steps.
4. **Removes deprecated `RemoteTransformerBlock`.**
`RemoteTransformerBlock` was deprecated a long time ago, before Petals 1.0.0. Its removal is long due.
5. **Removes `dht_utils.get_remote_module()`, `dht_utils.get_remote_sequence()`.**
This functions duplicate the functionality of the `RemoteSequential` constructor.
6. (minor) **Removes `RemoteSequential.is_subsequence` flag.**
This flag worked incorrectly and was never used. I am removing it for the sake of simplicity.
**Why?**
- We'd like to avoid excess threads for the original sequence manager in case if we only use its slices (e.g. when we add adapters or need only a subset of model blocks):
- If we create a sequence manager just before a fork (e.g. in a web app backend or a multi-thread benchmark), we'd like to avoid excess threads in the original process and only use this thread in child processes where we actually call `.make_sequence()`.
This PR makes servers return their free cache (in tokens * layers to make it compression-agnostic)
To be used when calling make_sequence(optimize="inference")
This pull-request implements a simple (1) greedy (2) latency-agnostic routing optimization that should speed up both our use cases.
Why this exists: our effort to merge full routing (ping-aware, throughut-aware, dijkstra) is in a sorry state between several branches; merging it into main would take many days.
Co-authored-by: Aleksandr Borzunov <borzunov.alexander@gmail.com>
This pull request adds an option to run Petals server on multiple local GPUs. It uses https://github.com/BlackSamorez/tensor_parallel
- 8bit approximation error same as in main (mean~=2% q0.9~=5%)
- TP=1, 2, 3 (see screenshots above)
- forward, grad w.r.t. input and inference exact match with main with TP=1
- `>=`80% GPU utilization with 3x 1080ti, batch = 8 tokens
- throughput measured with and without TP
- TP on 1080Tis has near-linear speedup comparable to the benchmarks (see first message)
Co-authored-by: Iaroslav Lisniak <yalisnyak@nes.ru>
Co-authored-by: Andrei Panferov <andrei@blacksamorez.ru>
Co-authored-by: Alexander Borzunov <borzunov.alexander@gmail.com>
1. Added `from petals.client import *` to `petals/__init__.py`, so you can write just that:
```python
from petals import DistributedBloomForCausalLM
```
I didn't do the same with server, since its classes are supposed to by used by `petals.cli.run_server`, not end-users. Though it's still possible to do `from petals.server.smth import smth` if necessary.
2. Fixed one more logging issue: log lines from hivemind were shown twice due to a bug in #156.
3. Removed unused `runtime.py`, since the server actually uses `hivemind.moe.Runtime`, and `runtime.py` has no significant changes comparing to it.
This pullrequest removes custom speed_test code in favour of speedtest-cli module.
This is necessary to ensure that random warnings / print-outs do not mess with our outputs.
Co-authored-by: Max Ryabinin <mryabinin0@gmail.com>
* Add missing methods for SamplingAlgorithm, fix docstrings
* Add SamplingAlgorithm to _choose_sample_algorithm
* Add test_sampling
* Add a warning if sampling options were passed, but do_sample=False
* Skip the sampling test for now
Co-authored-by: Alexander Borzunov <borzunov.alexander@gmail.com>
- latest accelerate, transformers, huggingface_hub
- rearrange attention caches to support https://github.com/huggingface/transformers/pull/18344
- remove unused code
- fix edge case where session crashes when receiving seq length 0
- assert transformer version when importing WrappedBloomBlock
Co-authored-by: Alexander Borzunov <borzunov.alexander@gmail.com>
Co-authored-by: Max Ryabinin <mryabinin0@gmail.com>
- Linear8bitLt now supports for pre-turing GPUs by temporarily upcasting quantized weights.
- added a test for linear8bitlt accuracy with the new fallback, the accuracy is similar than the real thing, (slightly better due to non-quantized A)
- performance is roughly halfway between the default mode and memory_efficient_backward
Alternatives considered:
- cupy - slow, casting to float internally
- triton - fast but unstable af. every 3rd attempt to matmul is a segfault
- bnb.functional.igemm (no lt) - "CuBLAS Error 8" on old GPUs
Co-authored-by: Aleksandr Borzunov <borzunov.alexander@gmail.com>
A patch to bitsandbytes 0.34.0 that introduces an option to run backward pass in default (fast) matrix layout.
Authors: cxb inversion by @borzunov, original 8bit code by @timdettmers
* optimized layout inversion code by @borzunov ([original code](https://colab.research.google.com/drive/1EJ0MKifajXSSVq7O2_QGwtb0l6gRAGrh?usp=sharing)) to use less forward calls
* implemented CustomLinear8bitLt, a child of Linear8bitLt that can do backward without CB
* added exact match tests for layouts and linear layers: see tests/test_linear8bitlt.py
* switched petals to the new layer type
Core idea: layouts apply the same permutation to every tile in the matrix. We can treat this as (batched) gather ops.
Reshape input tensor so that ij-th gather operation op will apply to ij-th elements in each tile.
Prototype:
Layout info: https://github.com/TimDettmers/bitsandbytes/blob/main/csrc/kernels.cu#L2130-L2136
Co-authored-by: Alexander Borzunov <hxrussia@gmail.com>
Co-authored-by: Aleksandr Borzunov <borzunov.alexander@gmail.com>
Co-authored-by: Tim Dettmers <tim.dettmers@gmail.com>
- [x] made RemoteSequenceManager into a background thread that pre-fetches information instead of running just in time
- [x] moved routing-related stuff to petals.client.routing
- [x] extract remote peer routing information to RemoteSequenceInfo
- [x] made sure that the code survives continued use (e.g. one hour)
- [x] updated every spot where update_ is called manually
- [x] modified get_sequence to check that the thread is alive, warn if not
- [x] removed max_retries, switched rpc_info to exponential backoff
- [x] fixed a bg that causes RemoteSeq* to lose user-defined hyperparameters (e.g. timeout) upon subsequencing (sequential[3:5])
- [x] moved client-side points strategy to client.routing
- [x] ensured that RemoteSequenceManager thread created in get_remote_module properly shuts down when the module is destroyed
- [x] resolved minor affected todos
- [x] modified tests to no longer use PYTHONPATH
- [x] worked around protocol error in rpc_info
Co-authored-by: Aleksandr Borzunov <borzunov.alexander@gmail.com>
Co-authored-by: Artem Chumachenko <artek.chumak@gmail.com>
Currently, the schemas use `torch.float32`, so all inputs and outputs converted to float32 before sending and after receiving on both servers and clients. This creates a huge slowdown for the system.
* This PR makes the schemas use the server's `--torch_dtype` argument (default is `torch.bloat16` for BLOOM-176B)
* an option for client to request a specific output compression. Use case 1: client sends quantized inputs and expects quantized inputs in return. Use case 2: client uses quantization for gradients w.r.t. activations, but keeps grads w.r.t. __prompts__ as is for greater precision.
* a comment explaining the purpose of NoSpendingPolicy - since we likely won't have it for the workshop
* a test with custom compression (janky implementation for testing purposes)
Co-authored-by: justheuristic <justheuristic@gmail.com>
1. Petals can be now installed using `pip install git+https://github.com/bigscience-workshop/petals`
- In case if you already cloned the repo, you can do `pip install .` or `pip install .[dev]`
2. Moved `src` => `src/petals`
- Replaced `from src.smth import smth` with `from petals.smth import smth`
3. Moved `cli` => `src/petals/cli`
- Replaced `python -m cli.run_smth` with `python -m petals.cli.run_smth` (all utilities are now available right after pip installation)
4. Moved the `requirements*.txt` contents to `setup.cfg` (`requirements.txt` for packages is not supported well by modern packaging utils)
5. Increased the package version from `0.2` to `1.0alpha1`
* priority in handlers and backend pools
* simple points system on server side
* priortize task in handler before submit task
* fix tests
* s/expert/block/g
Co-authored-by: justheuristic <justheuristic@gmail.com>
- Maximum length is now provided in `.inference_session(max_length=100)`
- previously, we would always assume max length = 2048
- added a generic way to forward **kwargs to inference session
- for compatibility with #47
- Note to @borzunov : it does *not* pass them arbitrarily, but instead checks for kwarg names at the bottom level
- run_server can be started with a custom max_length for inference
- renamed --cache_size_bytes to --attention_cache_bytes (to avoid collision with --cache_dir)
- --attn_cache_bytes can now support humane file sizes (e.g. 300MB instead of 314572800)
- made some server-side errors more human-readable to user (e.g. when max length is exceeded)
Co-authored-by: Aleksandr Borzunov <borzunov.alexander@gmail.com>
Co-authored-by: Alexander Borzunov <hxrussia@gmail.com>
Artem Chumachenko
Denis Mazur
Max Ryabinin
Alexander Borzunov
justheuristic
Pavel Samygin