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>
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>
Add --offload-towers: place multimodal tower weights in pinned host
memory and bulk-copy each unit to the GPU only during its forward,
keeping towers out of VRAM between encodes. Unlike UVA zero-copy,
whole-tensor H2D transfers avoid re-reading operand tiles over PCIe
inside the GEMM inner loop. Tower streaming is independent of layer
group scheduling, so group_size may be 0 for tower-only offload.
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>
Intake of open upstream PR vllm-project/vllm#58863. Active only with
VLLM_USE_V2_MODEL_RUNNER=1; inert otherwise.
Speculative decoding on Gated DeltaNet snapshotted the full recurrent
state after every draft position and reserved one Mamba block per draft
position. On a 24 GB card that is reserved VRAM we never use, plus
per-step copy traffic.
RecoverSSM, the protocol Kimi-K3 KDA already uses, replaces this:
verify runs from a read-only checkpoint and stores a small per-token
replay record [correction (V), normalized key (K), gate (1)] in fp32;
commit replays only the accepted tokens and writes the recurrent state
once, plus the block-boundary state in align mode. The PLE dilated
short conv compacts its accepted window with the same plan.
Conflict in vllm/models/qwen4_exp/nvidia/mtp.py: our get_top_tokens()
work put LocalArgmaxMixin on Qwen4ExpMTP, the PR adds SupportsReplaySSM
to the same base list. Kept both.
Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
Intake of open upstream PR vllm-project/vllm#58449.
With MTP num_speculative_tokens > 1, the V2 autoregressive speculator
reuses one attention-metadata build across draft steps only when every
group's builder sets supports_draft_decode_metadata_update. The QSA
side-cache builder (QWEN4_EXP_EXP_QSA_STATE) did not, so every draft
step rebuilt attention metadata for the whole model.
QSAMetadataBuilder now implements update_draft_decode_metadata: it
re-launches the kernel with in-place seq_lens updates, so one metadata
build serves every draft step.
Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
Intake of open upstream PR vllm-project/vllm#56742. Every rig here runs
MTP on both Qwen3.8 models, so these three bugs sit on our hot path.
The MTP hidden buffer was allocated with torch.empty and no device, so
it followed the ambient default instead of device_config.device; on a
CPU-default process the draft buffer lands on the wrong device. Both
NVIDIA and AMD backends now call make_mtp_hidden_buffer() in
vllm/models/qwen4_exp/common/mtp.py, which pins it to the configured
device.
Config parsing now accepts configs carrying M-RoPE keys and the
Transformers spelling of layer_types; both previously failed at load.
Triton warmup: Qwen4Exp was missing from the Qwen warmup dispatch, so
its Triton kernels compiled during the first served requests instead of
at startup.
Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
On a hybrid model (full-attention + Mamba/GDN layers, e.g. Qwen3.6-35B-A3B),
enabling MTP/EAGLE speculative decoding together with prefix caching corrupts
the recurrent state, and the server serves wrong answers. #43559 reports
tool-eval accuracy dropping from ~90% to ~50%. Nothing raises an exception
and nothing shows up in the logs, so an operator has no signal that anything
is wrong, and prefix caching then spreads it, because the corrupted state is
hash-reachable and every later request sharing that prefix restores from it.
The root cause is the find_longest_cache_hit method on MambaManager,
at vllm/v1/core/single_type_kv_cache_manager.py:1049 on main, which takes
drop_eagle_block in its signature and never acts on it. Its body searches
right-to-left for the last cached block, builds [null]*i + [cached], and
breaks, referencing neither drop_eagle_block nor .pop(.
Signed-off-by: MOVZX <movzx@yahoo.com>
use_local_argmax_reduction hard-fails at worker startup when the draft
model lacks get_top_tokens(): the V2 speculator probes for the method
in _validate_local_argmax_reduction and raises, with no fallback to
the full-vocab logits path. Qwen4ExpMTP (nvidia and amd variants) has
compute_logits but no get_top_tokens, so any Qwen4Exp MTP drafter with
the flag enabled crashed both TP ranks during load_model.
Qwen4ExpMTP owns a flat self.lm_head and its compute_logits is exactly
self.logits_processor(self.lm_head, hidden_states), the same shape as
Qwen3_5MTP, so LocalArgmaxMixin applies unchanged: vocab-parallel
argmax with a (value, index) gather, O(2*tp_size) vs O(vocab_size)
communication per draft token. No D2T remapping needed since the
draft vocabulary equals the target vocabulary.
Signed-off-by: MOVZX <movzx@yahoo.com>
Adds an opt-in backend, --engram-config '{"host_file_gather": true}', that
serves Qwen4Exp PLE rows from the checkpoint files through the page cache
instead of allocating the 47.7 GiB table on GPU or in pinned host memory.
- Qwen4ExpPLEFileGatherEmbedding allocates no table rows; load_weights keeps
the default loader's zero-copy safetensors views of each shard. At bind
(post-load hook) every view must be file-backed per /proc/self/maps (HF
cache symlinks work; heap/pinned/anonymous memory is rejected) and the
shards must cover the whole table.
- Each step, Qwen4ExpModelState.prepare_inputs computes the PLE n-gram ids,
copies them to pinned host memory, syncs the stream once, issues
POSIX_FADV_WILLNEED over the coalesced page ranges of the step's rows,
gathers the rows on the host, and copies them into a fixed device staging
buffer. The forward reads only the staging buffer, so FULL CUDA graph
capture is unaffected.
- Rows stay in the checkpoint dtype (FP8 table stays FP8, dequantized on the
GPU); dummy loads serve zeros; weight reload is rejected; takes precedence
over cpu_offload; rejected on ROCm and with elastic EP.
Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
Fixes#58807. A compressed-tensors pack-quantized checkpoint that ships
its MTP head dense (plain mtp.*.weight names in a separate bf16 shard)
crashes at weight load: Qwen3_5MultiTokenPredictor builds its Linears
from the target quant config, so the loader expects packed names
(weight_packed) the checkpoint never had, and the weight-map walk dies
with "no module or parameter named fc.weight".
Detect the dense MTP head from the checkpoint itself: read the
safetensors index via the same repo-file channel used for recipe.yaml
(local cache first, Hub download as fallback). If any mtp.* tensor
carries a quantization artifact suffix the head is properly quantized
and nothing changes; if every mtp.* entry is a plain .weight the whole
predictor builds unquantized (fc and layers), mirroring the existing
modelopt_fp4 and GPTQ dynamic carve-outs. Single-file checkpoints have
no index and keep current behavior; NVFP4 checkpoints are unaffected
(their mtp.layers.* carry artifacts).
Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
The pinned-host PLE lookup Triton kernel only gathers and stores bytes,
but passing an fp8e4nv tensor to it fails on architectures without
native FP8 pointer support (e.g. SM86) with an error saying type
fp8e4nv is not supported (supported fp8 dtypes are fp8e4b15, fp8e5).
Since FP8 E4M3 and int8 both occupy one byte per element, reinterpret
the PLE weight view and the output tensor as int8 when the weight dtype
is float8_e4m3fn; the non-FP8 path is unchanged.
Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
Apply the qwen3_5_quantized_embedding patch: construct the embedding
through the model's quant config so quantized (e.g. fp8/int8)
embeddings are supported, and give it a stable prefix for weight
loading.
Co-Authored-By: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
MOVZX
