dusterbloom: feat(bonsai-q1): packed 1.25-bpw engine with fp16 attention path#13
Open
dusterbloom wants to merge 13 commits intomainfrom
Open
dusterbloom: feat(bonsai-q1): packed 1.25-bpw engine with fp16 attention path#13dusterbloom wants to merge 13 commits intomainfrom
dusterbloom wants to merge 13 commits intomainfrom
Conversation
dusterbloom
force-pushed
the
dusterbloom/bonsai-q1-fp16
branch
2 times, most recently
from
e9a3152 to
39a657d
Compare
This was referenced May 4, 2026
fix(deps): update rust crate toml to v1
…l-action-digest chore(deps): update taiki-e/install-action digest to cca35ed
…file chore(deps): update rust crate tokio to v1.52.2
…-lockfile chore(deps): update rust crate tower-http to v0.6.9
…anbanda#143) Adds AnyCache::trim_by to roll back KV layers for speculative decode while leaving hybrid Arrays state untouched.\n\nCI: https://github.com/panbanda/higgs/actions/runs/25312580791
…#148) * feat(qwen3_next): mixed-bit Qwen3.5 GDN BA loading fallback Adds a fallback path for loading Qwen3.5 models with mixed-bit GDN projection weights (some layers q4, some q8 — common in unsloth's dynamic-quant variants). The default fused-projection loader fuses `in_proj_a` + `in_proj_b` into a single matmul; mixed-bit weights have incompatible shapes and the fusion fails. Behaviour: 1. Detect via `is_mixed_bit_gdn_ba_fusion_error` — matches a `ModelError::ShapeMismatch` whose message contains both `in_proj_ba` and `requires separate GDN projections`. 2. On detection, retry the load with `args.use_separate_gdn_projections = true`, taking the `load_qwen3_5_moe_weights_direct` path. Forward dispatches go from 2 to 4 GDN ops per layer — slightly slower but correct. 3. Forced separate (via `args.use_separate_gdn_projections` config or `HIGGS_SEPARATE_GDN_PROJ` env var) skips the fused attempt entirely. Also adds: * `qwen3_5_quantization_config` — parses `{group_size, bits}` from the per-layer `quantization` map in `config.json`. * `qwen3_5_mixed_ba_quantization_layers` — scans for the layers where `in_proj_a` and `in_proj_b` differ in bits or group_size. * `can_concatenate_axis0` — guard used inside `load_qwen3_5_moe_weights_fused` to emit the diagnostic `ShapeMismatch` error rather than panicking on the concat. * `load_qwen3_5_model_with_gdn_fallback` — private helper called by both `load_qwen3_5_model` (dense) and `load_qwen3_5_moe_model` (MoE), unifying the fallback path. Adaptations from feat/magic-canvas → origin/main: * The dense `load_qwen3_5_model` previously only honoured the env var; now it honours `args.use_separate_gdn_projections` too, matching the MoE path. Strict improvement: the config flag is set only by the env var or by mixed-bit detection. * No `unwrap()`, no `as` casts (use `i32::try_from`); match arms enumerate variants. No file-level allows added. Verification on origin/main (rustc 1.95.0): * `cargo check -p higgs-models` — clean * `cargo clippy --all-targets --all-features -- -D warnings` — clean * `cargo fmt --check` — clean * `cargo test -p higgs-models --lib` — 333/333 pass (3 new) Source: feat/magic-canvas commit `061e500c`. Direct cherry-pick had 5 conflict regions because origin/main has evolved the load functions independently; this is a manual surgical port that preserves origin/main's structure while adding the fallback behaviour. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(qwen3_next): preserve explicit GDN projection config --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-authored-by: Jonathan Reyes <me@jonathanreyes.com>
…panbanda#141) * perf(models): fused MoE gate+up — 3→2 expert matmuls per layer SwitchMlpWeights::forward_gather_fused() lazy-concatenates gate+up weights on first call, then dispatches a single gather_qmm instead of two separate calls. FfnBlock::forward() now routes through the fused path instead of forward_gather_global_sort(). Measured on 35B-A3B-3bit M4 base: - S=1 decode: 27ms → 17ms (−37%) - S=16 verify: 253ms → 112ms (−56%) - MoE/layer at K=1: 0.47ms (down from ~0.68ms) Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com> * style: cargo fmt qwen3_next.rs Reflow `let fw/fs/fb = ops::concatenate_axis(..)` from broken-line indentation back onto single lines so `cargo fmt --all -- --check` passes in CI. * fix(clippy): backtick MoE/gather_qmm doc + safe top_k u32 cast Two errors flagged by `-D clippy::doc-markdown` and `-D clippy::cast-sign-loss`/`-D clippy::as-conversions`: - Backtick `MoE` and `gather_qmm` in the `fused_gate_up` doc comment. - Replace `top_k as u32` with the same `u32::try_from(top_k).map_err(...)` pattern already used by `forward_gather_global_sort`. * fix(qwen3_next): gate MoE gate-up fusion behind opt-in --------- Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com> Co-authored-by: Jonathan Reyes <me@jonathanreyes.com>
panbanda
force-pushed
the
dusterbloom/bonsai-q1-fp16
branch
from
fad54d3 to
4af2603
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
2 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.
Summary
Adds Bonsai-Q1, a packed 1-bit Qwen3-shaped target model engine for 1.7B / 8B checkpoints. Routes via existing
model_type=\"qwen3\"arm: any checkpoint declaringquantization.bits == 1lands in the dedicated packed engine.Headline (production decode, M4 Max)
The win came from a dtype audit, not new architecture.
apply_yarn_ropewas multiplying fp16 q/k by an f32 scalar frommlx_rs::array!(mscale), silently upcasting the entire attention chain to f32. Bonsai's yarn factor=4.0 yields mscale≈1.139, so the branch fired every rope call (72×/step on 8B). Fix: cast the mscale scalar tox's dtype before multiply, keeping the chain in fp16.What lands
bonsai_q1::PackedQ1Linear— 1.25-bpw layer matching MLX's bits=1 QuantizedLinear / PrismML fork (1 bit/weight + 32-bit scale + bias per group of 128 input columns)bonsai_q1::BonsaiQ1Engine/BonsaiQ1Gpu— causal forward + KV cache on the packed engineyarn— shared YaRN rope helpers (apply_yarn_rope,compute_yarn_freqs,yarn_get_mscale) with the dtype-safe scalar castAnyModel::BonsaiQ1variant + match arms (forward, forward_batched, MTP, hidden_size, kv_cache_geometry, make_cache_with_config, image_size)is_bonsai_q1inmodel_loader— peeksconfig.jsonforquantization.bits == 1to route the qwen3 armSteppingKeyValueCache::key_value_arrays_mut— simultaneous mutable K/V access forUpdatablestate borrowingTest plan
cargo check -p higgs-models -p higgs-engine— cleancargo test -p higgs-models --lib -- --test-threads=1— 330 passed, 0 failed, 25 ignoredcargo test -p higgs-models yarn::(targeted) — yarn rope-dynamic ground-truth test passescargo run -p higgs-bench --release --bin bench_decode -- --model bonsai-8b-q1 --warmup 1 --trials 5 --max-tokens 200 --temperature 0— needs a Bonsai-8B-q1 entry inbenchmarks/models.tomlto land here; verify ≥60 tok/scargo clippy -p higgs-models -p higgs-engine— cleanNotes
yarn::tests::rope_dynamic_matches_static_offset_0_to_64) is#[ignore]d due to pre-existing harness flakiness — global MLX Metal/RNG state contamination from prior tests poisons the rope output. Passes targeted viacargo test -p higgs-models yarn::. The underlying behavior is what shipped the 2.83× win in production.