[0.10.2][Security] Remove structurally dead code (#444) (#460)

Cherry-pick #444 from branch releases/v0.11.0.

Signed-off-by: Artur Fierka <[email protected]>

Author: afierka-intel

vllm_gaudi/extension/bucketing/common.py

@@ -302,7 +302,6 @@ def get_filters(is_prompt, use_merged_prefill, use_contiguous_pa):
             return filters_map[phase][use_merged_prefill]
         else:
             return filters_map[phase][use_contiguous_pa]
-        return []
 
     buckets = set()
     buckets_2d = set()

vllm_gaudi/extension/ops.py

@@ -1041,12 +1041,6 @@ def scaled_fp8_quant(
     if scale is None:
         raise "dynamic scaled_fp8_quant not implemented for HPU"
         # TODO: calculate scale to match gaudi2 240 range instead of 448
-        if use_per_token_if_dynamic:
-            scale = torch.empty((input.numel() // input.shape[-1], 1), device=input.device, dtype=torch.float32)
-            torch.ops._C.dynamic_per_token_scaled_fp8_quant(output, input, scale, scale_ub)
-        else:
-            scale = torch.zeros(1, device=input.device, dtype=torch.float32)
-            torch.ops._C.dynamic_scaled_fp8_quant(output, input, scale)
     else:
         output = torch.ops.hpu.cast_to_fp8_v2(input, 1 / scale, False, False, dtype=torch.float8_e4m3fn)[0]
 

vllm_gaudi/v1/worker/hpu_model_runner.py

@@ -61,7 +61,6 @@
 from vllm.v1.worker.utils import (gather_mm_placeholders, sanity_check_mm_encoder_outputs, scatter_mm_placeholders)
 from vllm.v1.sample.rejection_sampler import RejectionSampler
 from vllm.v1.spec_decode.eagle import EagleProposer
-from vllm.v1.spec_decode.medusa import MedusaProposer
 from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
 from vllm.v1.spec_decode.ngram_proposer import NgramProposer
 from vllm.v1.sample.logits_processor import build_logitsprocs
@@ -668,7 +667,6 @@ def __init__(
                     self.use_aux_hidden_state_outputs = True
             elif self.speculative_config.method == "medusa":
                 raise NotImplementedError("Medusa speculative decoding is not supported on HPU.")
-                self.drafter = MedusaProposer(vllm_config=self.vllm_config, device=self.device)  # type: ignore
             else:
                 raise ValueError("Unknown speculative decoding method: "
                                  f"{self.speculative_config.method}")
@@ -3924,20 +3922,6 @@ def __del__(self):
     @torch.inference_mode()
     def profile_run(self) -> None:
         return
-        """Profile to measure peak memory during forward pass."""
-        # use an empty tensor instead of `None`` to force Dynamo to pass
-        # it by reference, rather by specializing on the value `None`.
-        # the `dtype` argument does not matter, and we use `float32` as
-        # a placeholder (it has wide hardware support).
-        # it is important to create tensors inside the loop, rather than
-        # multiplying the list, to avoid Dynamo from treating them as
-        # tensor aliasing.
-
-        # Run empty prefill forwards - prefill max batch and prefill max seq
-        self._execute_dummy_scenario((1, self.max_model_len, 0), None)
-        max_seq_len = math.ceil(
-            (self.max_num_tokens // self.max_prefill_batch_size) / self.block_size) * self.block_size
-        self._execute_dummy_scenario((self.max_prefill_batch_size, max_seq_len, 0), None)
 
     def _dummy_run(self, max_num_batched_tokens: int) -> None:
         assert max_num_batched_tokens == 1