fix preempted prompts (vllm-project#928)

The preempted prompts might failed to mitch the `num_computed_tokens <
num_prompt_tokens` test and be treated as decoding then cause runtime
error.

- add `_is_prompt()` to check if a request is prompt or not.
- consider the `num_scheduled_tokens` to handle the preempted prompts.
- add test for preemption handling to the CI.

---------

Signed-off-by: Youlei Yang <youlei.yang@intel.com>

Author: yangulei

tests/full_tests/ci_e2e_discoverable_tests.sh

@@ -351,6 +351,13 @@ run_mistral3_test() {
     echo "✅ Test with multimodal-support with Mistral-Small-3.1-24B passed."
 }
 
+# Preemption test
+run_preemption_test() {
+    echo "➡️ Testing preemption handling..."
+    VLLM_SKIP_WARMUP=true PT_HPU_LAZY_MODE=1 python -u "${VLLM_GAUDI_PREFIX}/tests/full_tests/preemption.py"
+    echo "✅ Test with preemption handling passed."
+}
+
 # Spec decode with ngram
 run_spec_decode_ngram_test() {
     echo "➡️ Testing Spec-decode with ngram..."

tests/full_tests/preemption.py

@@ -0,0 +1,43 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from vllm import LLM, SamplingParams
+
+# Sample prompts.
+prompts = [
+    "Hello, my name is",
+    "The president of the United States is",
+    "The capital of France is",
+    "The future of AI is",
+]
+# Create a sampling params object.
+sampling_params = SamplingParams(temperature=0.8, top_p=0.95, max_tokens=512, ignore_eos=True)
+
+
+def main():
+    # Create an LLM.
+    llm = LLM(
+        model="meta-llama/Meta-Llama-3-8B-Instruct",
+        block_size=128,
+        max_model_len=1024,
+        max_num_batched_tokens=1024,
+        gpu_memory_utilization=0.9,
+        num_gpu_blocks_override=8,  # to trigger preemption
+        disable_log_stats=False,
+    )
+    # Generate texts from the prompts.
+    # The output is a list of RequestOutput objects
+    # that contain the prompt, generated text, and other information.
+    outputs = llm.generate(prompts, sampling_params)
+    # Print the outputs.
+    print("\nGenerated Outputs:\n" + "-" * 60)
+    for output in outputs:
+        prompt = output.prompt
+        generated_text = output.outputs[0].text
+        print(f"Prompt:    {prompt!r}")
+        print(f"Output:    {generated_text!r}")
+        print("-" * 60)
+
+
+if __name__ == "__main__":
+    main()

vllm_gaudi/v1/worker/hpu_model_runner.py

@@ -1635,6 +1635,20 @@ def _get_num_decodes(self) -> int:
             num_decodes += 1
         return num_decodes
 
+    def _is_prompt(self, req_idx: int, scheduler_output: "SchedulerOutput") -> bool:
+        req_id = self.input_batch.req_ids[req_idx]
+        num_computed_tokens = int(self.input_batch.num_computed_tokens_cpu[req_idx])
+        num_prompt_tokens = int(self.input_batch.num_prompt_tokens[req_idx])
+        num_scheduled_tokens = scheduler_output.num_scheduled_tokens.get(req_id)
+        spec_decode_tokens = scheduler_output.scheduled_spec_decode_tokens.get(req_id)
+
+        num_decode_tokens = 1 if spec_decode_tokens is None else len(spec_decode_tokens) + 1
+        is_prompt = num_computed_tokens < num_prompt_tokens  # normal prompt
+        is_prompt = is_prompt or num_scheduled_tokens > num_decode_tokens  # maybe preempted prompt
+        is_prompt = is_prompt and not self.is_decoder_only(req_id)
+
+        return is_prompt
+
     def _get_prompts_and_decodes(
         self,
         scheduler_output: "SchedulerOutput",
@@ -1679,7 +1693,6 @@ def _get_prompts_and_decodes(
                         requests = metadata.reqs_to_store | metadata.reqs_to_load
             else:
                 requests = scheduler_output.kv_connector_metadata.requests
-
         # Traverse decodes first
         decode_req_ids = []
         num_computed_tokens_decode = []
@@ -1693,19 +1706,11 @@ def _get_prompts_and_decodes(
                         self.input_batch.req_type[req_id] = requests_type[req_id]
                         break
 
-            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[i]
-            num_prompt_tokens = self.input_batch.num_prompt_tokens[i]
-            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
-            if num_computed_tokens < num_prompt_tokens and \
-                not self.is_decoder_only(req_id):
-                # This is prompt
+            if self._is_prompt(i, scheduler_output):
                 break
 
-            # This is decode
-            # NOTE(chendi): To support spec decode,
-            # we don't assume num_scheduled_tokens == 1.
-
             decode_req_ids.append(req_id)
+            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[i]
             num_computed_tokens_decode.append(int(num_computed_tokens + 1))
 
         if self.profiler.enabled:
@@ -1718,16 +1723,12 @@ def _get_prompts_and_decodes(
             req_id = self.input_batch.req_ids[i]
             assert req_id is not None
 
-            num_computed_tokens = self.input_batch.num_computed_tokens_cpu[i]
-            num_prompt_tokens = self.input_batch.num_prompt_tokens[i]
-            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
-
             # Must be prompt
-            assert num_computed_tokens < num_prompt_tokens
-            # NOTE(kzawora): In preempted sequences, num_output_tokens can be > 0, and still be a valid prefill
+            assert self._is_prompt(i, scheduler_output)
 
             prompt_req_ids.append(req_id)
-            prompt_scheduled_tokens.append(num_scheduled_tokens)
+            num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
+            prompt_scheduled_tokens.append(int(num_scheduled_tokens))
 
         return PromptDecodeInfo(prompt_req_ids, decode_req_ids, prompt_scheduled_tokens)