[V1][TPU] Enable Top K

[NickLucche]

Enabling the topk optimization that was introduced in #15242.

Currently facing the very issue foreseen by @njhill here #15242 (comment).

ERROR 03-25 18:27:23 [core.py:343]     random_sampled = self.topk_topp_sampler(
ERROR 03-25 18:27:23 [core.py:343]                      ^^^^^^^^^^^^^^^^^^^^^^^
ERROR 03-25 18:27:23 [core.py:343]   File "/home/nick/vllm/.venv/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
ERROR 03-25 18:27:23 [core.py:343]     return self._call_impl(*args, **kwargs)
ERROR 03-25 18:27:23 [core.py:343]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 03-25 18:27:23 [core.py:343]   File "/home/nick/vllm/.venv/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
ERROR 03-25 18:27:23 [core.py:343]     return forward_call(*args, **kwargs)
ERROR 03-25 18:27:23 [core.py:343]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 03-25 18:27:23 [core.py:343]   File "/home/nick/vllm/vllm/v1/sample/ops/topk_topp_sampler.py", line 119, in forward_tpu
ERROR 03-25 18:27:23 [core.py:343]     topk_values, topk_indices = torch.topk(logits, k, dim=-1)
ERROR 03-25 18:27:23 [core.py:343]                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ERROR 03-25 18:27:23 [core.py:343] TypeError: topk(): argument 'k' (position 2) must be int, not Tensor

Dumping work for reference, will look into it asap.

Update:

• I've fixed topk only using @njhill work on [V1][Sampler] Faster top-k only implementation #15478 which should land soon on main
• Removed the env variable that would fallback to native impl (as dicussed in prev pr)
• Addressed some of the cruft leftover from [TPU][V1] Fix Sampler recompilation #15309

For completeness, I've run microbenchmarks and the new impl is slower (but of course correct):

//before
Running 32 elapsed time: 0.0018310546875
Running 32 elapsed time: 0.0017833709716796875
// after
 Running 32 elapsed time: 0.003275632858276367
Running 32 elapsed time: 0.003297090530395508

cc @hyeygit

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @NickLucche.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[NickLucche]

Let's hold until main is fixed to reduce entropy

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @NickLucche.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[hyeygit]

Thank you @NickLucche for this PR and thank you @njhill for fixing the batch case for top-k! In #15242 I only tested with the microbenchmark and test_sampler.py (scalar case only), not realizing that k can be batched. Thank you for the catch and sorry for the miss.

One thing to note is that on TPU torch.topk still involves a full vocab sort (see XLA lowering). The reason using torch.topk is so much faster on TPU is because of avoiding a full-vocab torch.scatter (as used in apply_top_k_top_p) which is extremely slow on TPU.

[NickLucche]

The main blocker for this PR is topk recompilation.
I was under the impression the XLA lowering would circumvent that need

import torch_xla.core.xla_model as xm
import torch_xla.debug.metrics as met
import torch

B, V = 3, 64
device = xm.xla_device()

logits = torch.randn(B, V, device=device)

# pre compile
k = 3
top_k_mask = logits.topk(k, dim=1).values
top_k_mask = top_k_mask.cpu()

print(met.short_metrics_report())
met.clear_all()

# Run
for k in [1, V]:
    top_k_mask = logits.topk(k, dim=1).values
    top_k_mask = top_k_mask.cpu()
    print(met.short_metrics_report()) # shows it compiles
    met.clear_all()

Is there something I am missing? @hyeygit @yaochengji

[hyeygit]

The main blocker for this PR is topk recompilation. I was under the impression the XLA lowering would circumvent that need

@NickLucche thank you for raising this. I think it's because the return shape of torch.topk is a function of k, so a change in k would cause XLA to recompile.

If this k-induced recompilation is not acceptable (sounds like it isn't), then let me implement a TPU-specific top-k using the same logic as in #15736. Will send out a PR shortly.

[NickLucche]

I think it's because the return shape of torch.topk is a function of k, so a change in k would cause XLA to recompile.

Yeah the reason is quite clear to me, I just don't know why I believed this wouldn't trigger recompilation lol

then let me implement a TPU-specific top-k using the same logic as in #15736.

That will do thanks a lot!
Then we have to put this PR on hold until the above gets merged or I cherry pick your work here and co-author.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @NickLucche.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[NickLucche]

@hyeygit I took the liberty of adding your contrib #15891 to this PR so we can test. Locally everything looks fine, let's wait for the CI tests. Nice work thanks!

[hyeygit]

So enabling topk here is costing us a bit of performance.

@NickLucche yes I think I can corroborate with this. I ran benchmark_serving.py with top-k and top-p enabled and observed a 6.8% drop in throughput (5.98 req/s -> 5.57 req/s). This was on TPU VM v6e-1 and running Llama3.1-8B.

[alexm-redhat]

Just ran this PR for llama 70B with 8 x v6e TPUs. It achieves 5 reqs/sec instead of the previous 5.1 reqs/sec so the performance penalty I see is negligible.

[yaochengji]

LGTM! Thanks, Nick!

[alexm-redhat]

LGTM! Just some small comments

[hyeygit]

@NickLucche small request -- could you incorporate the top-k equivalence test from my PR? https://github.com/vllm-project/vllm/pull/15891/files#diff-09d15417fe42d494c51aaa9635ad51536751cc3c0659a7c4ce3b66bd6900eb1f

[hgt312]

default top_k pad value should not be zero? 0 trilgget the error in probs_sort.gather(-1, top_k_count)