Skip to content

sync: open-core source for 0.4.2 #82

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
IsuraManchanayake merged 1 commit into from
Oct 27, 2026
Merged

sync: open-core source for 0.4.2 #82

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 8 additions & 8 deletions base-convert/Cargo.lock
View file
Open in desktop

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

2 changes: 1 addition & 1 deletion base-convert/Cargo.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@ members = [
]

[workspace.package]
version = "0.4.1"
version = "0.4.2"
edition = "2021"
license = "Apache-2.0"
repository = "https://github.com/basecompute/baseRT"
Expand Down
5 changes: 5 additions & 0 deletions base-convert/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,11 @@ basert pull basecompute/<name>
basert pull meta-llama/Llama-3.2-1B
basert pull Qwen/Qwen3-0.6B --target base-q8

# A variant by the name `basert list` gives it (`default-q8`, `cuda-q4mix`,
# a `file-…` variant), or a quant; the form `basert chat` and `serve` take:
basert pull basecompute/<name>:default-q8
basert pull Qwen/Qwen3-0.6B:q8

basert pull <id> --dry-run # resolve + print the plan, download nothing
basert list # installed models (table; --json for JSON)
basert list --remote # also show catalog models not yet installed
Expand Down
11 changes: 8 additions & 3 deletions base-convert/crates/base-arch/src/dflash.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,9 +12,14 @@
//! `DFlash2DraftModel`, `Qwen3DSparkModel`) — converting it as a Qwen3 would
//! drop `fc` / `hidden_norm` and write a broken LM.
//!
//! Tensor policy: every tensor stays f16 (the drafter's acceptance rate is
//! the whole point of the scheme, and the runtime's block step is bound by
//! the target's verify anyway). `fc.weight` is split by input column into
//! Tensor policy: f16 by default; under base-convert's `--quantize-drafter`
//! the transformer layers and the `fc.{i}` taps follow the target scheme
//! (`--keep-drafter-fc-f16` leaves the taps f16), while norms, the DFlash 2
//! selector codebooks and the Markov / confidence heads stay f16. The
//! catalog's default-q4 drafters quantize the layers; `basert pull` converts
//! a sidecar that way, the taps included for this family only, where 4-bit
//! taps were measured (DFlash2, 2026-10-19: acceptance unchanged), not for
//! DSpark or EAGLE-3. `fc.weight` is split by input column into
//! one `fc.{i}.weight` per tap (base-convert's `FcSplitProvider`), so the
//! runtime runs one GEMM per tap block of the tap-major input.
//!
Expand Down
Loading
Loading