Method profile_run

autoregressive/serve/model_runner.py:889–953

(self)

    @torch.inference_mode()
    def profile_run(self) -> None:
        # Enable top-k sampling to reflect the accurate memory usage.
        sampling_params = SamplingParams(top_p=0.99, top_k=self.vocab_size - 1)
        max_num_batched_tokens = self.scheduler_config.max_num_batched_tokens
        max_num_seqs = self.scheduler_config.max_num_seqs

        # This represents the maximum number of different requests
        # that will have unique loras, an therefore the max amount of memory
        # consumption create dummy lora request copies from the lora request
        # passed in, which contains a lora from the lora warmup path.
        dummy_lora_requests = []
        dummy_lora_requests_per_seq = []
        if self.lora_config:
            for idx in range(self.lora_config.max_loras):
                lora_id = idx + 1
                dummy_lora_request = LoRARequest(
                    lora_name=f"warmup_{lora_id}",
                    lora_int_id=lora_id,
                    lora_local_path="/not/a/real/path",
                )
                self.lora_manager.add_dummy_lora(dummy_lora_request,
                                                 rank=LORA_WARMUP_RANK)
                dummy_lora_requests.append(dummy_lora_request)
            dummy_lora_requests_per_seq = [
                dummy_lora_requests[idx % len(dummy_lora_requests)]
                for idx in range(max_num_seqs)
            ]

        # Profile memory usage with max_num_sequences sequences and the total
        # number of tokens equal to max_num_batched_tokens.
        seqs: List[SequenceGroupMetadata] = []
        # Additional GPU memory may be needed for vision encoding, which needs
        # to be accounted for when calculating the GPU blocks for
        # vLLM blocker manager.
        # To exercise the worst scenario for GPU memory consumption,
        # the number of seqs (batch_size) is chosen to maximize the number
        # of images processed.
        if self.vision_language_config:
            max_num_seqs = min(
                max_num_seqs,
                int(max_num_batched_tokens /
                    self.vision_language_config.image_feature_size))
        for group_id in range(max_num_seqs):
            seq_len = (max_num_batched_tokens // max_num_seqs +
                       (group_id < max_num_batched_tokens % max_num_seqs))
            seq_data, fake_multi_modal_input = _prepare_fake_inputs(
                seq_len, self.vision_language_config)
            seq = SequenceGroupMetadata(
                request_id=str(group_id),
                is_prompt=True,
                seq_data={group_id: seq_data},
                sampling_params=sampling_params,
                block_tables=None,
                lora_request=dummy_lora_requests_per_seq[group_id]
                if dummy_lora_requests_per_seq else None,
                multi_modal_data=fake_multi_modal_input,
            )
            seqs.append(seq)