Npu decode funcall gating - #254
Draft
cavusmustafa wants to merge 34 commits into
Draft
Conversation
…g_src to recorde the src ggml tensor for OpenVINO dynamic shape infer
enable qwen35 Fix after rebase remove logging
…t reason: the backend test initializes unary op inputs over a wide range, [-150, 150]. For FP32, exp(x) overflows around x ~= 88.7, so this test can randomly generate values right in or beyond the overflow region
In stateful mode the NEOX RoPE branch fed rank-3 data ([S, n_heads, head_size]) into the Multiply against the rank-4 cos/sin tables ([1, S, 1, n_dims/2]). That mixed-rank broadcast is miscomputed by the OpenVINO GPU plugin, corrupting the rotated Q/K and producing garbage output (e.g. Phi-3-mini). Lift the data to rank-4 before the split/ Multiply so the operands are equal-rank, matching what the TYPE_NORMAL branch already does. CPU and stateless paths are unaffected. Phi-3-mini-Q4_K_M, wiki.test perplexity, GPU stateful: before: PPL = 27120.43 after: PPL = 6.2263 (CPU reference: 6.2251)
…ov name in ov bk; 3) fix issue in arch test & op test with latest code update
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
…layer-embedding archs On NPU, the decode (1-token) model compiles the non-repeated blocks as NPUW "funcall-everywhere" subgraphs. For gemma3/4-style architectures those blocks carry the per-layer input embedding (inp_per_layer / per_layer_proj), so each one pays a per-call function prologue (tensor bind + closure unpack) on every decoded token. At batch-1 decode this per-call overhead dominates and makes decode ~8x slower than it should be (gemma4 E2B Q4_K_M: ~1.6 -> ~13 t/s). Fix: give the decode model its own compile config (compile_config_decode) and, for archs that carry a per-layer input embedding, set NPUW_FUNCALL_FOR_ALL=NO for the decode model only. Prefill is unchanged (funcalls there are amortized over the whole prompt). The arch is detected via a per_layer tensor in the graph (ModelParams::has_per_layer_embd); other archs and CPU/GPU are unaffected. Verified no change on Llama-3.2-1B (no per-layer embedding). Adds optional env overrides for experimentation: GGML_OPENVINO_NPUW_FOLD / _FUNCALL_FOR_ALL (both prefill+decode) GGML_OPENVINO_DECODE_NPUW_FOLD / _FUNCALL_FOR_ALL (decode only; precedence)
cavusmustafa
force-pushed
the
npu-decode-funcall-gating
branch
from
July 15, 2026 23:18
74ad6de to
18d9d77
Compare
wine99
force-pushed
the
dev_backend_openvino
branch
from
July 27, 2026 04:33
4250de7 to
175f422
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
openvino: NPU decode fix for gemma3/4 — disable NPUW funcall for per-layer-embedding archs
On NPU, the decode (1-token) model wraps its non-repeated blocks as NPUW funcall
subgraphs. For gemma3/4-style architectures those blocks hold the per-layer input
embedding (
inp_per_layer/per_layer_proj), so each pays a per-call functionprologue (tensor bind + closure unpack) on every decoded token. At batch-1 decode
this overhead dominates.
Fix: the decode model gets its own compile config, and for archs carrying a
per-layer input embedding we set
NPUW_FUNCALL_FOR_ALL=NOfor decode only.Prefill is unchanged; other archs and CPU/GPU are unaffected. Arch detected via a
per_layertensor in the graph (ModelParams::has_per_layer_embd).Results (NPU, c=1024, decode steady-state):
Trade-off: decode-model compile time rises for gated archs (paid once per
process); prefill and CPU/GPU are unaffected.
Adds optional env overrides for experimentation:
GGML_OPENVINO_[DECODE_]NPUW_FOLD/_FUNCALL_FOR_ALL.