The ghosts of vllm

Buried 2026-10-16 by Tyler Michael Smith in 9eb7fb915
“[Build] Remove unused `tools/install_deepgemm.sh` (#61265)”

The hands that crafted it now rest beneath the soil it remembers. A shadow lingers, a quiet echo in the remnants of what was.

• recipes/models/MiniMaxAI/MiniMax-M2.7.yaml:244 (config)[install_deepgemm.sh](https://github.com/vllm-project/vllm/blob/main/tools/install_deepgemm.sh)
• recipes/models/MiniMaxAI/MiniMax-M2.yaml:226 (config)[install_deepgemm.sh](https://github.com/vllm-project/vllm/blob/main/tools/install_deepgemm.sh)
• recipes/models/stepfun-ai/Step-3.5-Flash.yaml:183 (config)- **FP8 DeepGEMM:** For FP8, install DeepGEMM via [install_deepgemm.sh](https://github.com/vllm-project/vllm/blob/v0.16.0rc0/tools/install_deepgemm.sh).
• recipes/models/zai-org/GLM-5.1.yaml:160 (config)- **DeepGEMM (FP8, NVIDIA):** install via `install_deepgemm.sh` from vLLM repo
• recipes/models/zai-org/GLM-5.2.yaml:463 (config)- **FP8 performance:** DeepGEMM is required — install via `install_deepgemm.sh`.
• recipes/models/zai-org/GLM-5.3.yaml:479 (config)- **FP8 performance:** DeepGEMM is required — install via `install_deepgemm.sh`.
• and 3 more

Buried 2026-07-25 by Harry Mellor in 3e74c60b9
“[Docs] Use `gen-files` for generated docs content (#49587)”

It rests here now, silent and gray as the stones around us. A forgotten echo lingers in the lines of code, a ghost of what once was.

• docs/mkdocs/gen_files/generate_argparse.py:336 (code)"cli/complete.md": (create_parser(CompleteCommand.add_cli_args), False),

Buried 2026-07-25 by Harry Mellor in 3e74c60b9
“[Docs] Use `gen-files` for generated docs content (#49587)”

The code remembers its existence still a faded echo in the project’s heart. It rests here among the lost pieces, a quiet shadow waiting for completion.

• docs/mkdocs/gen_files/generate_argparse.py:337 (code)"cli/run-batch.md": (create_parser(run_batch.make_arg_parser), True),

Buried 2025-09-15 by Woosuk Kwon in 759ef49b1
“Remove V0 Encoder-Decoder Support (#24907)”

The tests still whisper its name, a faded echo in the code’s heart. It rests here now, a forgotten piece beneath the stones.

• .buildkite/scripts/hardware_ci/run-cpu-test-ppc64le.sh:40 (tests)# pytest -v -s tests/models/language/generation/test_bart.py -m cpu_model

Buried 2026-07-30 by Julien Denize in 1a20d23da
“[PARSER][Mistral] unified engine-based parser for reasoning and tool calls (#48947)”

The tests still whisper its name, a lingering echo in the code’s heart. It rests here now, alongside forgotten intentions.

• tests/parser/mistral/test_reasoning.py:61 (tests)`tests/reasoning/test_mistral_reasoning_parser.py`.

Buried 2025-09-24 by Thomas Parnell in 969b4da3a
“[V0 Deprecation] Remove placeholder attn (#25510)”

The space where it rested is cold now a hollow echo. Dust settles upon the memory of what once was.

• tests/plugins/vllm_add_dummy_platform/vllm_add_dummy_platform/dummy_attention_backend.py:4 (tests)from vllm.attention.backends.placeholder_attn import PlaceholderAttentionBackend

Buried 2026-07-01 by Harry Mellor in a22e0dfc6
“[Model] Remove AyaVision, MusicFlamingo (#47263)”

The stone remembers a shape long vanished from the earth. Dust settles on the echoes of its absence still.

• vllm/model_executor/models/cohere2_vision.py:3 (code)# Adapted from vllm/model_executor/models/aya_vision.py

Buried 2026-06-15 by Ben Browning in 76a373eff
“[Frontend] Replace legacy Gemma4 parsers with engine-based implementation (#45588)”

The echoes of those frontend systems still reach for it in the shadows. It rests here, a silent fragment of what once was.

• vllm/reasoning/gemma4_utils.py:13 (code)non-streaming), see ``vllm.reasoning.gemma4_reasoning_parser``.

Buried 2026-06-15 by Ben Browning in 76a373eff
“[Frontend] Replace legacy Gemma4 parsers with engine-based implementation (#45588)”

Here, in this forgotten corner of the codebase, it rests silent now. A ghost of intention lingers where it once was.

• vllm/tool_parsers/gemma4_utils.py:13 (code)non-streaming), see ``vllm.tool_parsers.gemma4_tool_parser``.

Buried 2024-09-23 by Alex Brooks in 9b8c8ba11
“[Core][Frontend] Support Passing Multimodal Processor Kwargs (#8657)”

Rest here, little fragment, lost to the shifting earth. I remember the echoes of your presence within these lines.

• vllm/models/deepseek_v4/common/mm_preprocess.py:7 (code)official repository's ``image_processor.py`` so that token counts bit-match
• vllm/models/deepseek_v4/common/mm_preprocess.py:265 (code)``image_processor.py`` (no ``auto_map`` processor), so this class wraps
• vllm/models/deepseek_v41/common/mm_preprocess.py:6 (code)official repository's ``image_processor.py`` so that token counts bit-match
• vllm/models/deepseek_v41/common/mm_preprocess.py:199 (code)``image_processor.py`` (no ``auto_map`` processor), so this class wraps

Buried 2025-09-20 by Woosuk Kwon in 7ed82d197
“[V0 Deprecation] Remove V0 MP executor (#25329)”

It has settled into the quiet earth for a long while now. The echoes of its absence linger still within these shadowed records.

• docs/design/multiprocessing.md:119 (docs)WARNING 12-11 14:50:37 multiproc_worker_utils.py:281] CUDA was previously
• docs/usage/troubleshooting.md:184 (docs)WARNING 12-11 14:50:37 multiproc_worker_utils.py:281] CUDA was previously

Buried 2026-03-10 by Chang Su in 507ddbe99
“feat(grpc): extract gRPC servicer into smg-grpc-servicer package, add --grpc flag to vllm serve (#36169)”

It rests here now, silent and gray, a forgotten echo of work long past. The earth claims it still, a presence lingering in the shadowed code.

• .gitignore:247 (code)# Generated gRPC protobuf files (compiled at build time from vllm_engine.proto)

Buried 2026-03-20 by Yongye Zhu in 87bd91892
“[MoE Refactor] Mxfp4 oracle rebased (#37128)”

The hands that crafted it now rest beneath the soil. It lingers still, a quiet echo in the shadowed code.

• vllm/model_executor/layers/fused_moe/experts/trtllm_mxfp4_moe.py:288 (code)Moved from trtllm_moe.py.