Conversation
Integrates main's "Target Naga IR directly (#435)", segment-anything (#432), unified conformance library (#434), and README rename. Conflict resolutions of note: - kalosm-language: kept ner's kalosm-chunking crate extraction; wired it into main's feature-gated layout (chunking feature now pulls dep:kalosm-chunking; dropped the now-unused direct srx dep, which lives in kalosm-chunking). - fusor-core/matmul: took main's restructured matmul; ported ner's batch-dim broadcasting (out_shape max + relaxed assertion) into matmul/kernel.rs. - fusor-core/tensor: ported ner's Tensor::materialized() and the wasm32 guard on debug_assert_real() into main's tensor/ module; added the matching gpu::Tensor::materialized() wrapper in fusor. - fusor/lib.rs: took main's unified F16/F32 q_mat_mul arm (dequantize+mat_mul); kept ner's Tensor::materialize()/materialized() additions. - fusor/layers/linear.rs + quantized.rs + activations.rs: took main's versions (q_mat_mul_3d, unified gelu with tanh_exact, tests moved to conformance crate) which supersede ner's equivalent fixes. - rbert self_attention: kept ner's flash_attention(BatchKeyMask) path, which matches the idiom main already uses in modern_bert and is supported by the merged fusor flash_attention API. - device.rs: kept main's rewrite + ner's FUSOR_DISABLE_SHADER_F16 env guard. - rgliner / rgliner-web: aligned with main's fusor cpu/gpu backend split (added cpu/gpu features forwarding to fusor + rbert); removed the obsolete kalosm-common/tokio feature (main's kalosm-common now loads synchronously via ureq and no longer exposes tokio); ungated with_cache. Dropped as superseded by main's rewrites (verify rgliner batched inference): - compute_graph/resolve.rs dead-node-pruning fix: main rewrote the resolver (single-target -> multi-target, fusion rules split into separate modules with their own dead-node handling) and added fusion_correctness conformance tests; ner's single-target fix is structurally incompatible (references self.target). - quantized/matmul/sgemv batch guard: main's tile-ir qmatmul dispatches a dedicated batch_size dimension for the single-row path, covered by the quantized_matmul_batched conformance test. Verified: cargo check passes for the full workspace, the wasm32 rgliner-web demo, and the conformance test suite.
# Conflicts: # Cargo.lock
# Conflicts: # Cargo.lock # fusor-ml/fusor/src/lib.rs
ealmloff
added
enhancement
|
Preview available at https://floneum.github.io/kalosm/pr-preview/pr-433/ |
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
1 participant
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.
Adds support for entity recognition and linking with some small models in gliner family