Training support

.github/workflows/main.yml

@@ -160,7 +160,7 @@ jobs:
           ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
           GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}
           FUSOR_CONFORMANCE_REQUIRE_GPU: "1"
-        run: cargo test --lib --bins --tests --examples --workspace --features kalosm/language,kalosm/sound,kalosm/vision,kalosm/remote,kalosm/scrape --exclude fusor-core --exclude rbert --exclude rwhisper
+        run: cargo test --lib --bins --tests --examples --workspace --features kalosm/language,kalosm/sound,kalosm/vision,kalosm/remote,kalosm/scrape --exclude rbert --exclude rwhisper
 
   doc:
     if: github.event.pull_request.draft == false

.gitignore

@@ -23,3 +23,4 @@ rust-analyzer
 /.claude
 **/.claude/
 profiles
+/fusor-ml/fusor/examples/data

Cargo.toml

@@ -44,7 +44,9 @@ members = [
 ]
 
 [patch.crates-io]
-wgpu = { git = "https://github.com/ealmloff/wgpu", branch = "yield-now" }
+# Local checkout of the wgpu fork while the mixed-precision cooperative-MMA
+# extension bakes (f16 operands with f32 accumulators in the MSL writer).
+wgpu = { path = "../wgpu-fusor/wgpu" }
 
 [workspace.dependencies]
 kalosm = { path = "./interfaces/kalosm", version = "0.4.0", default-features = false }
@@ -390,3 +392,7 @@ codegen-units = 1
 
 [profile.dev.package."*"]
 opt-level = 3
+
+[patch."https://github.com/ealmloff/wgpu"]
+naga = { path = "../wgpu-fusor/naga" }
+wgpu = { path = "../wgpu-fusor/wgpu" }

fusor-ml/conformance/goldens/attention_gqa_causal.txt

@@ -0,0 +1,4 @@
+out 0xa45dec1e98442bd9
+dq 0x9d3eefd367974641
+dk 0x712914ed2040211b
+dv 0xeedffc029ea157a5

fusor-ml/conformance/goldens/bilstm_trace.txt

@@ -0,0 +1 @@
+out 0xab42f2b1f7f2e4b7

fusor-ml/conformance/goldens/decode_trace_dispatch.txt

@@ -0,0 +1,5 @@
+399 [("flash_attention", 6), ("matmul_f32", 19), ("merged_matmul", 6), ("nary_direct", 33), ("row_program", 13)]
+1 [("flash_attention", 6), ("matmul_f32", 19), ("merged_matmul", 6), ("nary_direct", 32), ("row_program", 13)]
+1 [("flash_attention", 6), ("matmul_f32", 19), ("merged_matmul", 6), ("merged_region", 1), ("merged_row", 1), ("nary_direct", 35), ("row_program", 15)]
+1 [("flash_attention", 24), ("matmul_f32", 59), ("merged_matmul", 18), ("merged_region", 40), ("merged_row", 32), ("nary_direct", 114), ("row_program", 27)]
+1 [("flash_attention", 24), ("matmul_f32", 59), ("merged_matmul", 18), ("merged_region", 39), ("merged_row", 26), ("nary_direct", 114), ("row_program", 27)]

fusor-ml/conformance/goldens/kernel_bench_dispatch.txt

@@ -0,0 +1,9 @@
+wgrad merged_matmul 2
+wgrad256 merged_matmul 2
+wgrad256m merged_matmul 2
+fwd merged_matmul 2
+fwd256 merged_matmul 2
+fwdup merged_matmul 2
+attn merged_matmul 2
+softmax merged_row 1
+softmax row_program 1

fusor-ml/conformance/goldens/qgemv_decode_ggml.txt

@@ -0,0 +1,4 @@
+q4k_4096x8192 0x7f4870e80e75707a
+q4k_4096x8193 0x92adc0e985bb5224
+q4k_4096x5120 0x253507257e08480d
+q6k_4096x8192 0x5394b93db2f8ac10