PYSEC-2026-3542

Summary

A frontend-legal multi-request speculative workload can make vLLM produce an out-of-vocabulary recovered token equal to vocab_size, convert that value to -1 when choosing the next live token for a request, and then feed that -1 back into the next drafter input ids. On Qwen3 GPTQ this reaches the worker-side drafting / attention path and crashes the engine with a GPU device-side assert.

The same issue is reachable through the public gRPC request surface by sending a specific overlapping Generate / Abort sequence.

Impact

• A remote client that can send public gRPC generation requests can crash the shared vLLM engine worker
• The triggering request sequence aborts concurrent requests and prevents later requests from completing until the worker is restarted
• In shared deployments, this is a service-wide denial of service for other clients, not just a failure isolated to the attacking requests
• The failure is reproducible, so repeated request sequences can sustain the outage

Affected version

• Confirmed on vLLM 0.17.1
• Earlier and later versions have not been checked yet in this report

Repro model

• Official Hugging Face repo:
  • Qwen/Qwen3-0.6B-GPTQ-Int8
• Anyone wants to reproduce the bug with my PoC scripts should download Qwen3-0.6B-GPTQ-Int8 first

Trigger chain

1. A legal multi-request speculative workload keeps structured-output state, speculative decoding, overlap, and request cancellation active in the same live engine.
2. During rejection sampling, vLLM produces a recovered token equal to the model vocab_size boundary value.
3. That recovered token appears in position 0 of the sampled speculative row for a live request. The same row also contains trailing padding entries equal to -1, but those padding entries are not the key fault by themselves.
4. The next-token preparation step treats the position-0 recovered token as the real next token for that request and converts that out-of-vocabulary value to -1.
5. The drafter writes that converted -1 back into the live next-step input-id row for the request.
6. The drafting / embedding / attention path later consumes that live invalid token and the worker crashes on GPU.

Details

Simple example

The important distinction is:

• trailing -1 values in a speculative row can be ordinary padding
• the bug appears when the first live token for a request becomes 151936 == vocab_size, and that live token is then converted into -1

In simplified form, the bad transition looks like this:

sampled speculative row:
[151936, -1, -1, -1, ...]

At this point, the trailing -1 values are only padding. The critical problem is that the first position holds 151936, which is out of vocabulary and is being treated as the request's real next token.

Then vLLM prepares the next-token buffer:

next_token_ids:
[-1, ...]

Finally, that converted -1 is written back into the live model input ids:

input_ids_after:
[-1, 0, 0, 0, ...]

The crash happens because the live next token became -1 and was later consumed by the drafting / embedding / attention path, not merely because the speculative row contained padded -1 entries.

Trigger path in code

1. The workload is frontend-legal. The requests use normal SamplingParams features such as structured outputs, stop, bad_words, min_tokens, and streaming overlap. No malformed token-id list is required at the request boundary.
2. In speculative decoding, the rejection sampler can generate recovered tokens when drafted tokens are rejected.

   # vllm/v1/sample/rejection_sampler.py
   def sample_recovered_tokens(...):
       recovered_token_ids = torch.empty_like(draft_token_ids)
       sample_recovered_tokens_kernel[(batch_size, max_spec_len)](...)
       return recovered_token_ids
   

   On the verified Qwen3 run, the recovered-token trace shows recovered_token_ids[0] = 151936, which is exactly vocab_size for this checkpoint.
3. The speculative proposer then prepares the next-token row from the sampled speculative row.

   # vllm/v1/spec_decode/eagle.py
   def prepare_next_token_ids_padded(...):
       ...
       eagle_prepare_next_token_padded_kernel[grid](
           sampled_token_ids,
           discard_request_mask,
           backup_tokens_gpu,
           next_token_ids,
           valid_sampled_tokens_count,
           gpu_input_batch.vocab_size,
           ...
       )
       return next_token_ids, valid_sampled_tokens_count
   

   In the verified trace, this step receives a sampled row beginning with 151936, followed by -1 padding. The important point is that 151936 occupies the first live token position for the request. This step then produces next_token_ids[0] = -1, meaning the live next token for the request has been converted to -1.
4. The drafter then rotates the draft input ids and inserts those next_token_ids back into the live input-id buffer.

   # vllm/v1/spec_decode/eagle.py
   def set_inputs_first_pass(...):
       ...
       self.input_ids[token_indices_to_sample] = next_token_ids
   

   In the verified trace, this produces input_ids_after[0] = -1.
5. The model-side embed path later consumes those input ids.

   # vllm/model_executor/models/qwen2.py
   def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
       return self.embed_tokens(input_ids)
   

   In the verified trace, this is the first point where the converted -1 becomes visible as a real model input. The bug is not merely that the sampled speculative row contained padding -1; the bug is that the live next token for the request became -1 and was written back into input ids.
6. After that point, the visible sink depends on timing and backend state. On the attached Qwen3 reproducer, the engine commonly dies later in the drafting / attention path with CUDA error: device-side assert triggered, for example under flash_attn_varlen_func(...).

Local script breakdown

repro_g4_recovered_minus1_local.py is a standalone local reproducer.

• It reads the Qwen3 checkpoint path from VLLM_POC_G4_MODEL or the built-in /path/to/qwen3 placeholder
• It creates EngineCore directly without any external helper dependency
• It submits one fixed multi-request workload that preserves the same overlap and speculative-decoding state needed for the bug
• It writes:
  • request_payloads.json
  • repro_config.json
  • timeline.json
  • responses.json
  • error.txt
  • recovered_chain_trace.jsonl
• recovered_chain_trace.jsonl is the key attribution artifact. It records the recovered-token chain directly from the standalone reproducer

gRPC script breakdown

repro_g4_recovered_minus1_grpc.py is a standalone public gRPC reproducer.

• It reads the Qwen3 checkpoint path from VLLM_POC_G4_MODEL or the built-in /path/to/qwen3 placeholder
• It starts a temporary vllm.entrypoints.grpc_server process
• It sends only public Generate and Abort RPCs
• It submits one fixed overlapping request sequence that preserves the same speculative-decoding state needed for the bug
• After the crash window, it sends one more public Generate probe request to confirm that later gRPC requests also fail after the worker dies
• It writes:
  • request_payloads.json
  • timeline.json
  • server_command.json
  • responses.json
  • post_crash_probe.json
  • server.stdout.log
  • server.stderr.log

Observed result

Local repro typically ends with:

• a recovered-token trace showing:
  • sample_recovered_tokens_return -> recovered_token_ids[0] = 151936
  • prepare_next_token_ids_padded -> next_token_ids[0] = -1
  • set_inputs_first_pass -> input_ids_after[0] = -1
  • embed_input_ids_out_of_range -> input_ids[0] = -1
• CUDA error: device-side assert triggered
• a fatal engine-side failure

gRPC repro typically ends with:

• the triggering gRPC requests failing with INTERNAL: EngineCore encountered an issue. See stack trace (above) for the root cause.
• server logs showing the worker dies with CUDA error: device-side assert triggered
• a later public probe request also failing after the worker is dead

This demonstrates that the issue is reachable through the public gRPC request surface, not only through a local reproducer.

Log snippets

Local recovered-chain trace

sample_recovered_tokens_return:
  recovered_token_ids = [151936, ...]
  vocab_size = 151936

prepare_next_token_ids_padded:
  sampled_token_ids_head = [[151936, -1, -1, ...], ...]
  next_token_ids = [-1, ...]

set_inputs_first_pass:
  input_ids_after = [-1, 0, 0, 0, ...]

embed_input_ids_out_of_range:
  input_ids = [-1, 0, 0, 0, ...]

gRPC server log

torch.AcceleratorError: CUDA error: device-side assert triggered
...
vllm.v1.engine.exceptions.EngineDeadError: EngineCore encountered an issue. See stack trace (above) for the root cause.
...
Error in Generate for request post_crash_probe
vllm.v1.engine.exceptions.EngineDeadError: EngineCore encountered an issue. See stack trace (above) for the root cause.

Root cause

This is a speculative-decoding state-handling bug, not an invalid frontend token-id input bug.

The root cause is that a recovered speculative token can become equal to vocab_size, then be selected as the live next token for a request, then be converted to -1, and that converted -1 is still written back into live drafter input ids and later consumed by the drafting / embedding / attention path.

For the Qwen3 checkpoint used here:

• 151936 == vocab_size

This value should be described as the model vocab_size boundary value, not as a legal token id.

Attachments

The attached bundle for this report should contain:

• repro_g4_recovered_minus1_local.py
• repro_g4_recovered_minus1_grpc.py

These two standalone scripts are sufficient to reproduce the issue and its public gRPC reachability.

Fix

A fix for this vulnerability has been merged in: https://github.com/vllm-project/vllm/pull/44744