A symbol scan over every tracked file found 123 top-level definitions in
vllm/ with zero references, including references inside strings. All of
them came from features this fork deleted earlier.
- Delete dead attention helpers: MLADCPManager, split_prefill_chunks, the
reshape helpers, compute_varlen_chunk_metadata, the LINEAR backend's
unused BailingLinearAttentionBackend, and the sparse flash-attn entry
points.
- Delete the CPU and AMX op wrappers from _custom_ops.py: cpu_mla_decode,
cpu_mla_extend, cpu_mla_get_scheduler_metadata,
gather_mla_context_cache_cpu.
- Delete dead quantization helpers: awq_pack, gptq_pack,
gptq_quantize_weights, get_and_maybe_dequant_weights,
marlin_quant_fp8_torch, the rand_marlin_weight_*_like pair,
ocp_mx_weight_dtype_and_rows, get_fp8_block_weight_scale,
input_to_float8.
- Delete dead model helpers: resolve_visual_encoder_outputs,
run_dp_sharded_vision_model, fast_topk, get_layer_index,
is_pp_missing_parameter, skip_spec_layers, KimiK3MergedQKVGateLinear,
selective_scan_fn, causal_conv1d_update_torch.
- Delete dead utils: format_kib, format_mib, chunk_list, prev_power_of_2,
split_host_port, join_host_port, make_zmq_path, make_async_with_semaphore,
is_meta_module, is_non_overlapping_and_dense, supports_xpu_graph,
create_kv_caches_with_random*, and the unused has_* capability probes.
- Delete the unreachable deep-gemm mega-MoE warmup path, which imported
vllm.models.deepseek_v4.nvidia.model and vllm.models.kimi_k3.nvidia.model,
neither of which exists anymore.
- Drop the imports and the two now-empty guards left behind.
Verified statically: every changed file compiles, no unused import
remains, and no deleted name is referenced in vllm, tests or csrc. The
runtime gates still need the Python environment rebuilt.
Co-authored-by: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
Production runs one static model behind the OpenAI-compatible API. Every
adapter path and non-OpenAI HTTP surface is unreachable there.
- Delete the LoRA stack: vllm/lora, LoRARequest on the engine wire, the
scheduler, worker, cudagraph and MoE plumbing, LoRA metrics, the lora
CLI group, and the LoRA tests.
- Drop the plugin loader in vllm/plugins. Platform resolution is now
CUDA-only and the general/stat-logger plugin hooks are gone.
- Delete the speech-to-text and realtime entrypoints, the scale-out
render routes, and vllm/config/speech_to_text.py. GenerationTask is
generate-only.
- Cut DecodeBenchConnector and the disagg-prefill doc and image. The
Multi, Offloading and Example connectors stay; production uses the
CPU KV offload tier.
- Delete the ROCm-only fused KDA chunk kernel (2346 lines) with its CMake
arch gate, declarations and op registrations. The CUDA build never
defined the guard and no Python caller existed.
- Delete the orphan Mistral and Pixtral chat fixtures and examples/.
- Fix _check_model: removing the LoRA early return made every request
fail with "The model ... does not exist. Valid aliases: ...".
Verified on an RTX 3090: import sweep reports 1242 modules with 7 known
optional-dependency failures; a live server answers chat, completions,
responses, tokenize and detokenize with zero log errors and 404 for the
removed surfaces; collection yields 8558 tests with the 3 pre-existing
errors; ruff adds no new finding; the incremental _C build succeeds.
Co-authored-by: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
- Remove Anthropic, Cohere, SageMaker, decisions, structured-decisions,
rl, render-launcher and batch entrypoints; keep scale-out token routes
- Drop the Swagger static bundle and the /docs, /redoc routes
- Delete device-keyed tuning configs for fused_moe, quant utils, mamba
select-state and helion; no sm86 entry ever existed for any of them
- Remove ten unused rope variants plus the fbgemm_fp8 and fp_quant
quantizers and their config, metrics and loader plumbing
- Cut the CPU device build: csrc/cpu, cpu_extension.cmake, cpu
requirements and the setup.py cpu target path
- Delete 29 environment variables nothing reads and 14 chat-template
fallbacks for models this fork no longer registers
- Remove nine dead CUDA ops (deepseek_v4, kimi_k3, minimax, muse_spark,
dsv3 gemm) with their kernels, bindings, wrappers and CMake entries
- Delete the connector test suites that targeted removed connectors
- Cut the Mistral tokenizer and parser stack, its tests, and the
mistral_common and ijson dependency pins
- Restore the Qwen4Exp registry rows and nvidia/mtp.py that an earlier
cut deleted while keeping the model tree
Co-authored-by: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
- Remove P/D transfer connectors: nixl, nixl_chat, mooncake x2, hf3fs,
moriio, hisparse, example and hidden-states connectors
- Reduce weight_transfer to the base API used by model loading
- Remove kv_offload tiering, simple_kv_offload, ec cpu/example,
snapshot command, engine snapshot server
- Remove transformers model fallback, dsv4 tokenizers, dead platforms
cpu/tpu/xpu, flashmla third_party, dsa, megamoe/flydsl, dead attn
ops, dead layers (short_conv, fused_embed_norm, indexer_topk,
relu2 gemma_ar, lightning_attn, minimax_rms_norm), gemm_rs_ar
- Keep MoE LoRA op (imported eagerly by punica_gpu); drop fp8 variant
- Rebuild parser hubs around Qwen3 adapters only (36 dead adapters)
- Remove dead tests with the code they covered
Co-authored-by: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
- Drop dead model modules, keeping files live imports require
- Remove anthropic, cohere, codegen entrypoints and batch STT modes
- Delete unused attention backends, MoE experts, and NPU/XPU util leaves
- Strip dead third_party trees and ROCm leaves with no CUDA path
- Remove mooncake EC connectors: dependency missing since branch base
- Re-audit verified by full import sweep and Qwen3.5 smoke test
Co-authored-by: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
Checkpoints that ship calibrated k_scale/v_scale tensors (e.g.
unsloth/Qwen3.8-27B-NVFP4, which carries one pair per full-attention
layer) name them ...self_attn.{k,v}_scale, but the runtime parameters
live on the inner attention layer (...self_attn.attn.{k,v}_scale).
Qwen3_5Model.load_weights hands the iterable straight to
AutoWeightsLoader, which performs no scale-name remapping, so the
scales are silently dropped and --kv-cache-dtype fp8 runs at scale 1.0
with only a warning_once that fires even on success (the MTP draft
layer legitimately has no scales).
Route scale-suffixed names through maybe_remap_kv_scale_name first,
mirroring gemma4/gpt_oss. No behavior change for checkpoints without
scales.
Measured on RTX 5090 / Qwen3.8-27B NVFP4 / fp8 KV / MTP k=3 (GSM8K
greedy, n=300): 92.3% at scale 1.0 vs 93.7% with scales loaded
(bf16-KV control 94.5%).
Fixes#54623
Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
- Keep only Qwen3.5, Qwen3.5-MTP and Qwen4Exp model paths; trim the model registry to 41 entries
- Delete idle attention backends (MLA, FlashInfer, ROCm-only, quant backends) and dead GDN CuTeDSL kernels
- Remove inc/quark quant loaders; keep compressed-tensors and Marlin
- Drop the Rust frontend, pooling entrypoints, and idle test suites
- Rebuild kernels for sm_80;sm_86 and strip the extensions
- Verified: full import sweep clean (optional deps only) and a dummy-load Qwen3.5 engine run on RTX 3090 via the Triton GDN path
Co-authored-by: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
Qwen3_VisionTransformer derived its compute device from
patch_embed.proj.weight and fed pos_embed.weight straight into the
position-embedding interpolation kernel, so the last two anchor weights
had to stay in VRAM for a tower to be streamable at all.
Record the device the tower was built on before the offloader moves
anything, and fall back to it when no weight is on the accelerator.
Move the interpolation table to that device per encode, which costs one
8.4 MiB host-to-device copy for a 456 MiB tower. Towers now leave no
parameters in VRAM between encodes; only small buffers such as the fused
input norm and rotary inv_freq remain.
Also let a tower restrict streaming through tower_offload_children, and
warn when a tower resolves its device property to another device type,
so a tower that reads an anchor weight reports the cause instead of a
Triton pointer error.
Co-authored-by: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
With VLLM_ABLITERATION set, boot crashed: dynamo traces
abliteration.register from Qwen3NextGatedDeltaNet.__init__, and
buf.copy_(r) fails on the meta device during the compile-time model
construction pass. Disable tracing for register and guard the copy so
an unmaterializable buffer turns projection off with a warning.
Log activation at registration instead of inside project_: project_
runs in the compiled region, and a logger call there aborts dynamo.
The width guard uses the same escape hatch.
Also project the embedding that Qwen3NextMultiTokenPredictor builds
for itself, which the first pass missed, and add tests for the
2560-value Flash-Next direction against a 5120 model and for a
half-width activation.
Co-authored-by: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
orcarouter/Qwen3.8-27B-Uncensored removes refusal by editing weights. The
Swift release we serve is AutoRound W8A16, so re-editing weights means
re-quantizing. Project one unit direction r out of every residual-stream
writer output instead: x <- x - r (r . x), which equals the weight edit
W' = W - r(r^T W) on every writer feeding that stream (Arditi et al. 2024).
Sites: attention output for both the full-attention and GDN branches, the
MLP output, and token embeddings, in Qwen3NextDecoderLayer.forward and
Qwen3NextModel.embed_input_ids, which Qwen3_5DecoderLayer and Qwen3_5Model
inherit but whose base constructors they skip, so both register there too.
The MTP drafter reuses Qwen3_5DecoderLayer and calls its own
embed_input_ids, so the draft path is covered at the same width, and
_maybe_share_embeddings at PP=1 hands the draft the target embedding
module. Sequence parallelism shards dim 0 only, so the last dim is
hidden_size at every site.
VLLM_ABLITERATION names a JSON file with key refusal_direction, a
safetensors tensor, or a bare name resolved under
/opt/qwen38/abliteration then /data/abliteration. Empty or 0 is off.
register() always creates the buffer _abliteration_r, zeros when off, so
one compiled graph serves ON and OFF and toggling the switch
(declared in envs.py, hence in compile_factors) invalidates the compile
cache. project_() is an in-place rank-1 update with no full-size
temporary, and skips tensors whose width is not hidden_size.
Co-authored-by: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
MOVZXandHinata
