Named entity recognition and linking with gliner

.gitignore

@@ -20,6 +20,7 @@ tokenizer.json
 out.txt
 todo.md
 rust-analyzer
+*.gguf
 /.claude
 **/.claude/
 profiles

Cargo.toml

@@ -16,6 +16,7 @@ bench = false
 [workspace]
 members = [
     "models/rbert",
+    "models/rgliner",
     "models/kalosm-llama",
     "models/kalosm-tokenizer",
     "models/rwhisper",
@@ -29,6 +30,7 @@ members = [
     "interfaces/kalosm-streams",
     "interfaces/kalosm-parse-macro",
     "interfaces/kalosm-common",
+    "interfaces/kalosm-chunking",
     "interfaces/kalosm-model-types",
     "fusor-ml/core",
     "fusor-ml/gguf",
@@ -51,13 +53,15 @@ kalosm = { path = "./interfaces/kalosm", version = "0.4.0", default-features = f
 kalosm-sample = { path = "./interfaces/kalosm-sample", version = "0.4.0" }
 kalosm-parse-macro = { path = "./interfaces/kalosm-parse-macro", version = "0.4.0" }
 kalosm-common = { path = "./interfaces/kalosm-common", version = "0.4.0" }
+kalosm-chunking = { path = "./interfaces/kalosm-chunking", version = "0.4.0" }
 kalosm-model-types = { path = "./interfaces/kalosm-model-types", version = "0.4.0" }
 kalosm-language-model = { path = "./interfaces/language-model", version = "0.4.0" }
 kalosm-streams = { path = "./interfaces/kalosm-streams", version = "0.4.0" }
 kalosm-language = { path = "./interfaces/kalosm-language", version = "0.4.0", default-features = false }
 kalosm-sound = { path = "./interfaces/kalosm-sound", version = "0.4.0", default-features = false }
 kalosm-vision = { path = "./interfaces/kalosm-vision", version = "0.4.0", default-features = false }
 rbert = { path = "./models/rbert", version = "0.4.0", default-features = false }
+rgliner = { path = "./models/rgliner", version = "0.4.0" }
 kalosm-llama = { path = "./models/kalosm-llama", version = "0.4.0", default-features = false }
 kalosm-tokenizer = { path = "./models/kalosm-tokenizer", version = "0.4.0" }
 rwhisper = { path = "./models/rwhisper", version = "0.4.0", default-features = false }

fusor-ml/core/src/compute_graph/resolve/run.rs

@@ -533,7 +533,18 @@ impl Resolver {
             let encode_start = host_trace.then(Instant::now);
             let mut dispatch_index = 0usize;
             let mut command_index = 0usize;
-            let dispatches_per_pass = dispatches_per_pass(total_kernels);
+            // Bound the cumulative workgroups recorded into a single compute pass by
+            // the device's real per-dimension workgroup limit. Each
+            // `begin_compute_pass` is a GPU command-encoder; recording too many
+            // passes into one command buffer loses the device on Metal (a
+            // long-sequence encoder + BiLSTM resolves to ~2200 dispatches, and one
+            // pass per dispatch — the old policy for large graphs — meant ~2200
+            // passes, which faults; the same dispatches grouped into a handful of
+            // passes run fine). Grouping by the device-reported workgroup limit
+            // keeps the pass count at roughly `total_workgroups / limit`: a few
+            // dozen for the largest graphs, one for decode/prefill — no threshold.
+            let pass_workgroup_budget =
+                u64::from(device.limits().max_compute_workgroups_per_dimension);
             while command_index < commands.len() {
                 match &commands[command_index] {
                     CommandRecord::CopyBuffer(copy) => {
@@ -577,14 +588,20 @@ impl Resolver {
                                 label: Some("Resolver Direct Kernels"),
                                 timestamp_writes: None,
                             });
-                        let mut pass_dispatches = 0usize;
+                        let mut pass_workgroups = 0u64;
                         while command_index < commands.len() {
-                            if pass_dispatches >= dispatches_per_pass {
-                                break;
-                            }
                             let CommandRecord::Dispatch(record) = &commands[command_index] else {
                                 break;
                             };
+                            let dispatch_workgroups = record.dispatch.workgroup_count();
+                            // Always record at least one dispatch per pass; otherwise
+                            // close the pass before its cumulative workgroups would
+                            // exceed the device budget.
+                            if pass_workgroups > 0
+                                && pass_workgroups + dispatch_workgroups > pass_workgroup_budget
+                            {
+                                break;
+                            }
                             if let Some((query_set, _, _, _)) = &query_resources {
                                 pass.write_timestamp(query_set, (dispatch_index * 2) as u32);
                             }
@@ -596,7 +613,7 @@ impl Resolver {
                             }
                             dispatch_index += 1;
                             command_index += 1;
-                            pass_dispatches += 1;
+                            pass_workgroups += dispatch_workgroups;
                         }
                     }
                 }
@@ -715,14 +732,3 @@ fn direct_plan_binding_buffers(inputs: &[MirValue]) -> Vec<Vec<std::sync::Arc<wg
         .collect();
     vec![buffers]
 }
-
-fn dispatches_per_pass(total_kernels: usize) -> usize {
-    if let Ok(value) = std::env::var("FUSOR_RESOLVE_DISPATCHES_PER_PASS")
-        && let Ok(parsed) = value.parse::<usize>()
-        && parsed > 0
-    {
-        return parsed;
-    }
-
-    if total_kernels >= 1024 { 1 } else { usize::MAX }
-}

fusor-ml/cpu/benches/comparison.rs

@@ -1,6 +1,7 @@
 use candle_core::{Device, Tensor as CandleTensor};
-use criterion::{BenchmarkId, Criterion, Throughput, black_box, criterion_group, criterion_main};
+use criterion::{BenchmarkId, Criterion, Throughput, criterion_group, criterion_main};
 use fusor_cpu::__private::TypedTensor;
+use std::hint::black_box;
 
 const SIZES: &[usize] = &[64, 256];
 

fusor-ml/cpu/benches/fusion.rs

@@ -5,8 +5,9 @@
 //! memory multiple times (once per operation).
 
 use candle_core::{Device, Tensor as CandleTensor};
-use criterion::{BenchmarkId, Criterion, Throughput, black_box, criterion_group, criterion_main};
+use criterion::{BenchmarkId, Criterion, Throughput, criterion_group, criterion_main};
 use fusor_cpu::__private::{ConcreteTensor, TypedTensor};
+use std::hint::black_box;
 
 /// Benchmark fused operations (single memory pass) vs separate operations (multiple passes)
 fn bench_fusion(c: &mut Criterion) {

fusor-ml/cpu/benches/qmatmul.rs

@@ -1,8 +1,9 @@
 use aligned_vec::AVec;
-use criterion::{BenchmarkId, Criterion, Throughput, black_box, criterion_group, criterion_main};
+use criterion::{BenchmarkId, Criterion, Throughput, criterion_group, criterion_main};
 use fusor_cpu::__private::{ConcreteTensor, QuantizedTensor, TypedTensor};
 use fusor_cpu::BlockQ8_0;
 use half::f16;
+use std::hint::black_box;
 
 /// Helper to create a Q8_0 block from scale and data
 fn make_q8_0_block(scale: f32, data: [i8; 32]) -> BlockQ8_0 {

fusor-ml/cpu/benches/tensor_ops.rs

@@ -1,5 +1,6 @@
-use criterion::{BenchmarkId, Criterion, Throughput, black_box, criterion_group, criterion_main};
+use criterion::{BenchmarkId, Criterion, Throughput, criterion_group, criterion_main};
 use fusor_cpu::__private::{Abs, Add, ConcreteTensor, Div, Mul, Neg, Sqrt, Sub, TensorBacking};
+use std::hint::black_box;
 
 const SIZES: &[usize] = &[64, 256];
 

fusor-ml/fusor/src/composite/mod.rs

@@ -17,6 +17,7 @@ mod reductions;
 mod rope;
 mod shape;
 mod to_vec;
+mod transformer;
 mod upsample;
 mod where_cond;
 
@@ -25,3 +26,8 @@ pub use rope::{RopeCache, base_inverse_frequency};
 pub(crate) use shape::broadcast_shapes;
 pub use shape::{arange, arange_step, cat, stack};
 pub use to_vec::{ToVec, ToVec1, ToVec2, ToVec3};
+pub use transformer::{
+    AttentionBias, AttentionVariant, FeedForwardVariant, GatedActivation, GroupedAttention,
+    LlamaFeedForward, Norm, PhiFeedForward, RopeLike, SeparateAttention, TransformerBlock,
+    forward_attention_qkv_f32,
+};