Repository navigation
ggml-openvino: select the NPU token_embd type with GGML_OPENVINO_NPU_TOKEN_EMBD_QUANT - #359
Merged
Merged
Conversation
github-actions
Bot
added
documentation
Author
|
Hi @cavusmustafa @wine99 Can you help take a review for this? Thanks. |
ravi9
force-pushed
the
dev_backend_openvino
branch
from
October 24, 2026 06:11
c5cd082 to
947be59
Compare
…table 1. Optimize gather to fetch row before dequant 2. add NPU token_embd quantization option
…TOKEN_EMBD_QUANT GET_ROWS had two rewrites that gather the rows of a quantized table before dequantizing them: gather_compressed_rows, which runs on every device, and gather_before_dequant, enabled on NPU by GGML_OPENVINO_GATHER_BEFORE_DEQUANT. Both match the int4 and int8 chains built by make_int4_weights() and make_int8_weights(). Keep gather_compressed_rows and drop the second rewrite and its switch. The NPU type of a Q6_K token_embd.weight was set by two booleans, GGML_OPENVINO_TOKEN_EMBD_I8 and GGML_OPENVINO_TOKEN_EMBD_I4, and by GGML_OPENVINO_NPU_TOKEN_EMBD_QUANT. Keep only the string variable, with the values f16 (default), q8_0_c, q4_0_128 and q4_0_c, so the int8 and group-128 int4 settings stay available. An unknown value aborts, as for GGML_OPENVINO_NPU_REQUANT_POLICY, instead of silently using f16. The variable replaces the two booleans in the NPU compiled-model cache key.
Owner
|
FYI: I've resolved the conflicts and am running some basic model validation. Will update this PR soon. |
ravi9
force-pushed
the
dev_backend_openvino
branch
from
October 26, 2026 21:41
674af4a to
9e01c2f
Compare
ravi9
changed the title
Owner
|
Rebased onto the updated dev_backend_openvino. dev already gathers the packed rows before dequantizing ( |
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
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.
Overview
On NPU, a quantized token_embd is tied to the output projection. It was dequantized in full (about 1.3 GB in f16) on every token, and NPUW then passed the dequantized table between submodels. Decode was slower than with an F16 table.
We do an optimization GET_ROWS now gathers the rows of the quantized constants and their scales first, and dequantizes only those rows. Grouped and channel-wise int4/int8 layouts are handled, with or without zero points. Any other shape falls back to the plain Gather.
Additional information
Switch: GGML_OPENVINO_GATHER_BEFORE_DEQUANT is tri-state. When unset it is on for NPU and off elsewhere, a positive value forces it on, and 0 forces it off.
ggml-openvino-extra.cpp: a new NPU-only option, GGML_OPENVINO_NPU_TOKEN_EMBD_QUANT, with q4_0_128, q4_0_c, or F16 when unset. It applies to a Q6_K token_embd.weight and logs the chosen type once.
ggml-cpu base: llama-perplexity.exe -m Qwen_Qwen3.5-4B-Q4_0.gguf -f wiki.test.raw -c 512 --chunks 20 -dev none -ngl 0 --kl-divergence-base kld-base-wiki.text.raw.chunks20-512.bin
npu base: unset GGML_OPENVINO_GDN_PRESCALE, kld-wiki.text.raw.chunks20-512-base.log
With an F16 table the gather switch changes neither accuracy nor speed. So, we keep the GGML_OPENVINO_GATHER_BEFORE_DEQUANT on by default.
Requirements
Yes, The minimal case passed.
AI help analysis and profile.