[V1] TPU - Remove self.kv_caches #14309
Conversation
alexm-redhat
requested review from
WoosukKwon,
comaniac,
njhill,
robertgshaw2-redhat and
ywang96
as code owners
|
👋 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 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 🚀 |
There was a problem hiding this comment.
Please revert this - we should make a examples/offline_inference/tpu/ folder to keep this
| forward_context[layer_name].kv_cache = [kv_cache] | ||
|
|
||
|
|
||
| class ModelWrapperV1(nn.Module): |
There was a problem hiding this comment.
Is it possible to implement ModelWrapperV1 like this?
class ModelWrapperV1(nn.Module):
def __init__(self, model: nn.Module, num_kv_heads, num_blocks, block_size):
super().__init__()
self.model = model
self.num_kv_heads = num_kv_heads
...
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
inputs_embeds: Optional[torch.Tensor] = None,
is_profile_run: bool,
) -> torch.Tensor:
if not is_profile_run:
num_kv_heads = self.num_kv_heads
...
class TPUModelRunner:
def _dummy_run(
self,
num_tokens: int,
is_profile_run: bool,
) -> None:
self.model.forward(..., is_profile_run=is_profile_run)
There was a problem hiding this comment.
@heheda12345 this is not possible because num_blocks is not known until determine_num_available_blocks is done and initialize_kv_cache is executed.
There was a problem hiding this comment.
Can we pass a fake value first and update it after determine_num_available_blocks?
|
@heheda12345 the is_profile_run arg is set to False by default, so it should not be necessary to pass this parameter explicitly to set_forward_context(..) function in other backends. Is there any specific code example where you would need to specify it explicitly? |
|
If we add |
|
@heheda12345 I see makes sense. There is a new PR from Google that removes the problematic non-profile reshuffling #14310 which will affect this PR. I would wait for this PR to land first and then rebase. |
| self._dummy_run(num_tokens=num_tokens, is_profile_run=True) | ||
|
|
||
| # This is used after KV cache init | ||
| def dummy_run( |
There was a problem hiding this comment.
_dummy_run and dummy_run look confusing, do we really need this overload?
| self.model = model | ||
| self.kv_cache_shape = None | ||
|
|
||
| def set_kv_cache_shape(self, kv_cache_shape): |
There was a problem hiding this comment.
nit: we can probably get away without setters as long as we keep the class and the logic lean
| if kv_cache_shape_prev is None: | ||
| kv_cache_shape_prev = kv_cache_shape | ||
| else: | ||
| assert kv_cache_shape == kv_cache_shape_prev |
There was a problem hiding this comment.
qq: is this for ruling out some model architecture?
| forward_context = self.vllm_config.compilation_config \ | ||
| .static_forward_context | ||
| for layer_name, kv_cache in kv_caches.items(): | ||
| # NOTE: Use list because of v0 PP virtual engine. | ||
| forward_context[layer_name].kv_cache = [kv_cache] |
There was a problem hiding this comment.
nit: do you see any use in having this bit factored as a util, similarly to bind_kv_cache? We could re-use at least in tpu_worker
|
This pull request has merge conflicts that must be resolved before it can be |
|
This pull request has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this pull request should remain open. Thank you! |
|
This pull request has been automatically closed due to inactivity. Please feel free to reopen if you intend to continue working on it. Thank you! |
This PR removes self.kv_caches from the tpu_model_runner.py in V1, so that @heheda12345 #14098 can cleanly land.
@mgoin @NickLucche feel free to make a pass.