Repository navigation
DSv4 static cache - #46384
DSv4 static cache#46384IlyasMoutawwakil wants to merge 21 commits into
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
| ): | ||
| self._last_compile_config = compile_config | ||
| mark_repeated_decoder_layers_as_compile_regions(self) | ||
| install_compile_patches() |
There was a problem hiding this comment.
many patches need to land in torch before this can work natively, in torch 2.13 maybe, but for now torch.compiler.nested_compile_region doesn't seem worth it (the compile speedup is also not that crazy)
|
The diff adds two new classes, |
ArthurZucker
left a comment
ArthurZucker
left a comment
There was a problem hiding this comment.
Nice! Can you spend some time on the doc so its more human! Otherwise LGTM, we can default the max compress lenght by dividing max_new_tokens by the compression rate, probably already done?
| # ``"compressor"`` entry, at ``head_dim`` for both the rolling source-token buffer | ||
| # and the compressed-entry storage. CSA adds an ``"indexer"`` entry with different | ||
| # widths (see :class:`DeepseekV4CSAStaticCache` which overrides these tables). | ||
| self._buffer_feat_by_name: dict[str, int] = {"compressor": config.head_dim} |
There was a problem hiding this comment.
let's find better names!
| self.buffer_kv: dict[str, torch.Tensor | None] = dict.fromkeys(self._buffer_feat_by_name) | ||
| self.buffer_gate: dict[str, torch.Tensor | None] = dict.fromkeys(self._buffer_feat_by_name) |
There was a problem hiding this comment.
this is a bitweird / looks wasteful
| """Static analogue of :meth:`DeepseekV4HCACache.store_compression_weights`. | ||
|
|
||
| Decode (``new_n == 1``): shifts the rolling buffer left by 1, appends the new | ||
| token at the rightmost slot, returns the full ``(B, compress_rate, head_dim)`` | ||
| buffer — always one window's worth, even when the window isn't yet full | ||
| (left slots are zeros). The compressor emits one candidate entry; the | ||
| attention mask hides it until the window actually closes. | ||
|
|
||
| Prefill (``new_n > 1``): re-uses the dynamic logic — concat buffer's real tail | ||
| with the new kv, peel off the window-aligned prefix, and refresh the rolling | ||
| buffer with the last ``compress_rate`` tokens of the combined stream. | ||
| """ |
There was a problem hiding this comment.
let's make this human redable / human made :) :)
| return prior_kv, prior_gate | ||
|
|
||
|
|
||
| class DeepseekV4HCAStaticCache(StaticSlidingWindowLayer): |
There was a problem hiding this comment.
we might want to add to generation_deepseek_v4.py its not core modeling!
| r"""Static counterpart to :class:`DeepseekV4CSACache`: adds the ``"indexer"`` name | ||
| and per-name overlap state (Ca slice of the previous window's chunk) on top of | ||
| the HCA static buffers. | ||
|
|
||
| Overlap state is fixed-shape ``(B, compress_rate, head_dim)`` per name and starts | ||
| as zeros — the existing consumer ``new_kv[:, 0, :ratio] = prior_kv`` is a write, | ||
| not an add, so handing back zeros on the first call is observationally identical | ||
| to the dynamic cache's ``None`` (the slot was already zeros from ``new_zeros``). | ||
|
|
||
| Note: no ``layer_type`` is set on this class — see :class:`DeepseekV4HCAStaticCache`. |
There was a problem hiding this comment.
same let's put effort in making this human like !
ArthurZucker
left a comment
ArthurZucker
left a comment
There was a problem hiding this comment.
Nice! can you put the benches here as well! 🤗
There was a problem hiding this comment.
it's more practical for testing than a long string
|
[For maintainers] Suggested jobs to run (before merge) run-slow: deepseek_v4 |
What does this PR do?
probably still ugly, but works, still needs benching. made with claude's help.
Code Agent Policy
The Transformers repo is currently being overwhelmed by a large number of PRs and issue comments written by
code agents. We are currently bottlenecked by our ability to review and respond to them. As a result,
we ask that new users do not submit pure code agent PRs at this time.
You may use code agents in drafting or to help you diagnose issues. We'd also ask autonomous "OpenClaw"-like agents
not to open any PRs or issues for the moment.
PRs that appear to be fully agent-written will probably be closed without review, and we may block users who do this
repeatedly or maliciously.
This is a rapidly-evolving situation that's causing significant shockwaves in the open-source community. As a result,
this policy is likely to be updated regularly in the near future. For more information, please read
CONTRIBUTING.md.Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.