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>
The value is chosen as some safe value below average at https://health.petals.dev/
Note that if a server uses relays, the effective throughput will be further divided by 2 (see #399).
* 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
Since `petals.ml` DNS record is still unavailable, we're switching everything to https://petals.dev
Co-authored-by: Aleksandr Borzunov <hxrussia@gmail.com>
This PR:
1. **Adds shortest path routing for inference.** We build a graph with client-server and server-server latencies and compute costs, as well as empirically measured overheads. For client-server latencies, we ping possible first and last servers in a sequence in `SequenceManager.update()`. We penalize servers who may not have enough cache for our request. This uses info added to DHT in #355, #356, #358.
2. **Makes a server ping neighboring servers in addition to next ones.** This is to get an opportunity to change the server even before we use all its blocks (e.g., because a neighboring server is faster). This feature is not enabled though, since it increases graph size for N servers to O(N^2) - but we may enable it if needed.
3. **Fixes a `SequenceManager` bug with the first `update()`.** Previously, this update was likely to produce incorrect information and cause to `MissingBlocksErrors` until the next update happens.
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.
Currently, each `TransformerBackend.inference_step` looks for adapters and sets the correct adapter type for each block. This is not very expensive, but it can measurably affect inference time.
This pull request uses faster adapter switching with just one variable assignment, without iterating over block.modules().
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>
Before this PR, `free_disk_space_for()` was able to remove **(a)** only entire cached revisions (= git commits/branches) and **(b)** only from the repository we're loading right now.
This PR allows this functions to remove arbitrary files separately from any repositories.
This is useful for transition to Petals 1.2.0+, since it now uses original repos instead of the ones with converted models (see #323). In particular, the cache for `bigscience/bloom-petals` is now deprecated and should be removed in favor of `bigscience/bloom`. This is also useful as a way to free space before loading LoRA adapters (#335).
This PR:
- Adds benchmark scripts for inference, forward pass, and full training step (e.g. used for experiments in our paper).
- Fixes bug with dtypes in `petals.DistributedBloomForSequenceClassification`.
- (minor refactor) Moves `DTYPE_MAP` to `petals.constants` as a useful constant.
This PR adds `petals.AutoDistributed{Model, ModelForCausalLM, ModelForSequenceClassification}` classes, similar to their `transformers.Auto{Model, ModelForCausalLM, ModelForSequenceClassification}` counterparts.
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).
Now, if a user passes unexpected kwargs to `.generate()`, they are __ignored__ and the code continues working as if the argument was correctly supported. For example, people often tried passing `repetition_penalty` and didn't notice that it does not have any effect. This PR fixes this problem.
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.
- After #285, `load_pretrained_block()` uses `accelerate.utils.set_module_tensor_to_device()`
- In accelerate>=0.16.0, it saves the tensor in the dtype previously used by the model instead of dtype of the weights (https://github.com/huggingface/accelerate/pull/920)
- Because of that, blocks and attention caches used float32, which caused OOMs
- This PR makes `load_pretrained_block()` respect `torch_dtype` (default: `"auto"`, which means reading `torch_dtype` from `config.json`)
**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()`.
`use_auto_relay=True` makes the libp2p daemon look for relays to become reachable if we are behind NAT/firewall. However, being reachable is not necessary for the Petals client, and we should not spend the relays' capacity on this.
This PR fixes issues of #290:
- hivemind bfloat16 codec crashed on dummy tensors (with 0 elements), see https://github.com/learning-at-home/hivemind/pull/560 (this PR makes Petals depend on the latest hivemind version from the repo, it's temporary)
- transformers version check mismatched with the version allowed in `setup.cfg`
Also:
- This PR enables 8-bit by default for TP. Even though TP in 8-bit may be slower, we currently prefer to host more blocks to increase the network's stability.
- new bitsandbytes supports newer *and* older GPUs
- new hivemind supports a better bfloat16 codec
Co-authored-by: Alexander Borzunov <borzunov.alexander@gmail.com>
For some reasons, right now 15 sec is not enough to connect to the bootstrap peers in the public swarm, as reported by multiple users and observed by me. Increasing it to 120 sec until we find the root cause of the issue.
This PR increases `request_timeout`, since the previous default of 30 sec is not enough for many use cases.
Previously, we kept the request timeout low since we assumed that the server could freeze on dial if the target peer is behind a firewall. However, apparently, it won't freeze because libp2p has its own [dial timeout](https://github.com/libp2p/go-libp2p/blob/v0.26.0/core/network/context.go#L11).
Before this PR, `model.generate()` returned one excess token when resuming generation with an existing (the last token of the previous session, `session.last_token_id`). This is an unexpected behavior not convenient for the downstream apps, so this PR changes it until it's too late.
Even if the swarm seems to have at least 2 servers for each block, turning off on one of the servers could break it. That's because once a server is turned off, others may move to a better position, creating a significant downtime on their way. This PR prohibits switching blocks if it would make the swarm disjoint along the way.
This PR:
1. Shows the current Petals version and checks for updates on startup.
2. Reports the current version and DHT mode in `rpc_info()`, so it can be shown on http://health.petals.ml or used on clients for efficient routing.
Servers joining from behind NATs/firewalls usually take several minutes to join a libp2p relay before they become accessible from the outside Internet. Moreover, requests to such servers are slower and more likely to fail (e.g., if the server switches a relay at the moment). If such servers host certain DHT keys, the swarm may occasionally lose read/write access to these keys, which results in:
- Clients being unable to find any servers hosting a certain block.
- All servers starting rebalancing to the same place to close the alleged "gap" in the swarm.
This PRs modifies servers so that DHT keys are only hosted on **directly reachable** servers (the ones who aren't behind NAT/firewall). This way, DHT becomes more stable and works faster. Of course, trhe servers behind NATs/firewalls still accept requests for running inference/forward/backward for blocks they hold (it's more acceptable for this kind of requests to be slower or fail).
Co-authored-by: Alexander Borzunov <borzunov.alexander@gmail.com>
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>
If all servers holding a certain block are blacklisted, we should display errors from them instead of raising `No peers holding blocks`.
Indeed, if the error is client-caused, the client should learn its reason from the latest error messages. In turn, if the error is server/network-caused and we only have a few servers, we'd better know the error instead of banning all the servers and making the user think that no servers are available.
* Don't count open fds since it leads to AccessDenied crashes on some machines
* Use --load_in_8bit=False by default in case of tensor parallelism
* Install petals from PyPI in fine-tuning tutorials
- Added relay options to servers
- Enabled relay options by default
- Changed hivemind version to 1.1.5
- Moved reachability check to be performed after blocks are loaded
Co-authored-by: Alexander Borzunov <borzunov.alexander@gmail.com>