[Kernel][Backend][Model] Blocksparse flash attention kernel and Phi-3-Small model

[linxihui]

• Supports of Microsoft Phi-3-Small-8K and Phi-3-Small-128K models, which use blocksparse flash attention
• Prefilling Triton kernel for block-sparse attn
• Modified paged attention CUDA with the block-sparse attention, which allows hybrid sparsity pattern for each attention head.
• Use torch SPDA in prefilling phase for V100 or older GPUs, as well as CPU

This is joint work between Microsoft GenAI @linxihui, @beagleski, and vLLM @zhuohan123, @simon-mo @youkaichao.

[simon-mo]

I made a pass. I think once this PR adds unit test for both the Triton and PagedAttention kernels it should be good to go. You might also need to run clang-format to fix the merge conflict.

[linxihui]

I made a pass. I think once this PR adds unit test for both the Triton and PagedAttention kernels it should be good to go. You might also need to run clang-format to fix the merge conflict.

Thanks @simon-mo for the review. I'll add the mising unitests today.

[simon-mo]

I have tested the PR locally as well.

[AllenDou]

Phi-3-small's SPECIAL_TOKENS('<|******|>') will cause guided_grammar crash

  File "/root/vllm/vllm/engine/async_llm_engine.py", line 39, in _raise_exception_on_finish
    task.result()
  File "/root/vllm/vllm/engine/async_llm_engine.py", line 517, in run_engine_loop
    has_requests_in_progress = await asyncio.wait_for(
  File "/usr/lib/python3.10/asyncio/tasks.py", line 445, in wait_for
    return fut.result()
  File "/root/vllm/vllm/engine/async_llm_engine.py", line 491, in engine_step
    request_outputs = await self.engine.step_async()
  File "/root/vllm/vllm/engine/async_llm_engine.py", line 225, in step_async
    output = await self.model_executor.execute_model_async(
  File "/root/vllm/vllm/executor/gpu_executor.py", line 117, in execute_model_async
    output = await make_async(self.driver_worker.execute_model
  File "/usr/lib/python3.10/concurrent/futures/thread.py", line 58, in run
    result = self.fn(*self.args, **self.kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "/root/vllm/vllm/worker/worker.py", line 272, in execute_model
    output = self.model_runner.execute_model(seq_group_metadata_list,
  File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "/root/vllm/vllm/worker/model_runner.py", line 709, in execute_model
    logits = self.model.compute_logits(hidden_states, sampling_metadata)
  File "/root/vllm/vllm/model_executor/models/phi3_small.py", line 403, in compute_logits
    logits = self.logits_processor(self.lm_head.weight, hidden_states,
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1532, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1541, in _call_impl
    return forward_call(*args, **kwargs)
  File "/root/vllm/vllm/model_executor/layers/logits_processor.py", line 58, in forward
    logits = _apply_logits_processors(logits, sampling_metadata)
  File "/root/vllm/vllm/model_executor/layers/logits_processor.py", line 115, in _apply_logits_processors
    logits_row = logits_processor(past_tokens_ids,
  File "/root/vllm/vllm/model_executor/guided_decoding/outlines_logits_processors.py", line 53, in __call__
    allowed_tokens = self.fsm.allowed_token_ids(self.fsm_state[seq_id])
  File "/usr/local/lib/python3.10/dist-packages/outlines/fsm/fsm.py", line 329, in allowed_token_ids
    self.regex_fsm = RegexFSM(regex_string, self.tokenizer)
  File "/usr/local/lib/python3.10/dist-packages/outlines/fsm/fsm.py", line 123, in __init__
    regex_string, tuple(sorted(tokenizer.vocabulary.items()))
TypeError: '<' not supported between instances of 'str' and 'bytes'

server:
python3 -m vllm.entrypoints.openai.api_server --model microsoft/Phi-3-small-8k-instruct --tensor-parallel-size 1 --served-model-name modelx --disable-log-stats --trust-remote-code

client:

curl http://localhost:8000/v1/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "modelx",
        "prompt": ["Generate a sql state that select col_1 from table_1 where it is equals to 1"],
        "max_tokens": 20,
        "temperature": 0,
        "guided_grammar": "start: select_statement\r\nselect_statement: \"SELECT\" column \"from\" table \"where\" condition\r\ncolumn: \"col_1\" | \"col_2\"\r\ntable: \"table_1\" | \"table_2\"\r\ncondition: column \"=\" number\r\nnumber: \"1\" | \"2\""
    }'

#5068 add a test case.