[Bugfix][V1] Warm up slot-mapping kernel through BlockTable

[lesj0610]

Problem

The old V1 BlockTable slot-mapping Triton kernel can be first touched after startup warmup, when the JIT monitor is already active.

The normal model dummy/profile runs do not directly exercise BlockTable.compute_slot_mapping() for every runtime-specialized request shape. When the first real request hits an uncovered shape, _compute_slot_mapping_kernel can still compile during inference.

Approach

• Add BlockTable.warmup_compute_slot_mapping() and MultiGroupBlockTable.warmup_compute_slot_mapping() so the warmup goes through the same V1 block-table wrapper that runtime input preparation uses.
• Cover the slot-mapping specialization classes observed for single-token, short non-divisible, and block-sized request shapes while staying within the allocated block-table bounds.
• Call the V1 slot-mapping warmup from GPU worker initialization after the existing kernel warmup and before CUDA graph capture / JIT monitor activation.
• Keep the change scoped to the old V1 BlockTable slot-mapping path. V2 GPU block-table and KV-zeroing warmup paths are not changed here.

This keeps the warmup near the owner of the slot-mapping kernel contract instead of adding a broader synthetic kernel launcher.

Test Plan

python -m py_compile \
  vllm/v1/worker/block_table.py \
  vllm/v1/worker/gpu_worker.py \
  tests/v1/worker/test_block_table_warmup.py

python -m ruff check \
  vllm/v1/worker/block_table.py \
  vllm/v1/worker/gpu_worker.py \
  tests/v1/worker/test_block_table_warmup.py

python -m ruff format --check \
  vllm/v1/worker/block_table.py \
  vllm/v1/worker/gpu_worker.py \
  tests/v1/worker/test_block_table_warmup.py

python -m pytest tests/v1/worker/test_block_table_warmup.py -q
git diff --check

Test Result

• py_compile: passed
• ruff check: passed
• ruff format --check: passed
• tests/v1/worker/test_block_table_warmup.py: 2 passed
• git diff --check: passed

GPU runtime smoke was not rerun for this standalone PR, so absence of runtime JIT warnings remains to be validated on a CUDA serving run.

Checklist

• Purpose
• Test plan and results
• AI assistance disclosed

AI assistance: Codex.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.