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>
With --use-replayssm and speculative decoding, Qwen4Exp's Gated DeltaNet layers verify the
draft window from a read-only checkpoint and keep a per-token replay record
[correction(V), normalized key(K), gate(1)] in fp32. After sampling, one commit replays the
accepted tokens and writes the recurrent state once (plus the block-boundary state in align
mode), instead of snapshotting the state after every draft position. The PLE dilated short
conv follows the same protocol (accepted-suffix compaction). num_speculative_blocks = 0,
so no per-draft Mamba blocks are reserved.
- GDN verify/commit kernels: model_executor/layers/mamba/gdn/recoverssm_gdn.py
- metadata/backends: v1/attention/backends/gdn_recoverssm.py, ple_recoverssm.py
- the verify reads strided q/k/v views of the conv output and writes in place
- builders declare AttentionCGSupport.NEVER: FULL_AND_PIECEWISE falls back to PIECEWISE;
the fallback now also counts breakable CUDA graphs (empty splitting_ops, with or
without torch.compile) as piecewise
attention instead of dropping to NONE
- the generic Mamba metadata builder skips the Mamba2 ReplaySSM ring-buffer workspaces
when --use-replayssm selects RecoverSSM (the PLE short conv uses that builder)
- tests: GDN verify/commit vs the native kernels (fp32 and bf16 state, align boundaries,
mixed and ragged batches, extreme gates), PLE vs token-by-token decode, config selection
Signed-off-by: Jürgen Schmied <juergenschmied70@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
RecoverSSM's commit-plan and conv-window compaction kernels are model independent.
Move them from models/kimi_k3 into model_executor/layers/mamba/ops/recoverssm.py
(Kimi-K3 keeps its private names as import aliases; kernel source unchanged), add
mamba/recoverssm_utils.py (activation predicate, argument validation) and the
generic spec-row/request mapping helpers to v1/attention/backends/recoverssm_metadata.py.
CacheConfig.use_kda_recoverssm becomes use_recoverssm: the switch now selects RecoverSSM
for any model whose recurrent layers implement it, not only KDA.
No functional change for Kimi-K3.
Signed-off-by: Jürgen Schmied <juergenschmied70@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>