Dashboard: multi-model support (Stories110M + Qwen3-0.6B) with GQA-aware
text generation and KV cache. Weights & Biases logging (--wandb flag) for
loss, timing, power, and checkpoint events. Top-k=50 sampling to eliminate
garbage tokens from untrained vocab entries. Tokenizer reads any vocab size.
train.m: only save checkpoint when loss improves (best_loss tracking).
Implement Grouped-Query Attention (16q/8kv heads, head_dim=128) for
Qwen3-0.6B (28 layers, 596M params). Model configs moved to
models/*.h headers selected at build time via make MODEL=xxx.
Key changes:
- GQA-aware MIL kernels: sdpaFwd split from woFwd (Q_DIM!=DIM),
qBwd/kvBwd split from qkvBwd (different IC dimensions)
- K/V tile (KV_HEADS→HEADS) before SDPA backward, reduce after
- 10 kernels total, all model-agnostic via compile-time defines
- Makefile: make MODEL=qwen3_06b (default) or MODEL=stories110m
- Both models verified: Stories110M ~115ms/step, Qwen3 ~412ms/step
Three bugs prevented loss from converging below 5.5 (unigram plateau):
1. FP16 underflow in ANE backward matmuls: gradient (~8e-5) × weight (~0.036)
products flushed to zero in fp16. Fixed with global loss scaling (256×)
applied once to dlogits, divided out before Adam update.
2. Backward weight staging used raw weights instead of transposed — all 4
backward kernels (wotBwd, qkvBwd, ffnBwdW2t, ffnBwdW13t) now use
pre-transposed buffers (Wot_buf, Wqt_buf, etc.).
3. Added AdamW (decoupled weight decay, wd=0.1 for weights, 0.0 for norms),
activation clipping (act_clip=20), gradient clipping, cosine LR schedule,
per-layer IOSurface weight pre-staging, and vocab compaction.
Loss now drops 9.14 → 5.74 in 500 steps from random init (87ms/step).
Benchmark report now includes full Stories110M model configuration
(arch, layers, dims, kernels). README updated: 12-layer results
replace stale single-layer numbers, limitations reflect current state.
Community-submitted results for M1 Pro/Max, M3 Pro, M4 Pro/Max, M5.
Includes training performance, peak throughput, MIL compatibility
matrix, and structured JSON data.
- Vectorize adam_update with vDSP batch ops (stories_cpu_ops.h)
Replaces scalar per-element loop with vDSP_vsmul/vsma/vsq/vdiv
Expected ~3-4x faster for 2.4M parameter updates
- Vectorize model_adam_step ADAM_UPDATE macro with vDSP (backward.h)
Same batch ops pattern for the train.m model pipeline
- Replace cpu_accum_dW with cblas_sgemm (backward.h)
dW += dy^T @ x is a standard BLAS GEMM operation
Expected 5-10x faster for weight gradient accumulation
- Replace cpu_matmul_backward_dx with cblas_sgemm (backward.h)
dx = dy @ W^T is also a standard BLAS GEMM
- Add -framework Accelerate to train target (Makefile)
- Validate all fread() return values in model_load_weights (model.h)
- Check ane_eval() return values in ane_conv_eval (forward.h) and ane_eval_k (tiny_train.m)
- Log error details on ANE eval failure (ane_runtime.h)
- Thread-safe RMSNorm: replace global g_rms_tmp with local allocation (stories_cpu_ops.h)
- Bounds-check token indices in cross_entropy_loss, embed_lookup, embed_backward
- Atomic checkpoint writes via tmp+rename pattern (tiny_train.m)
- Non-destructive recompile: compile new kernels first, swap only on success (model.h)
- Validate fread() in load_checkpoint (tiny_train.m)
[MLModel compileModelAtURL:] fails on macOS 26, breaking inmem_bench,
sram_bench, and sram_probe. This switches all three to generate MIL text
and weight blobs programmatically in memory (matching the working
inmem_peak.m approach), bypassing CoreML disk compilation entirely.
- inmem_bench.m: replace CoreML compile + file read with genMIL/buildWeightBlob
- sram_bench.m: switch from _ANEClient/_ANEModel to _ANEInMemoryModel API
- sram_probe.m: same _ANEClient → _ANEInMemoryModel conversion
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Parse static pipeline JSON step/batch/perf lines for real-time updates
- Running elapsed time, ms/step from wall-clock timestamps, steps/sec
- Compute ANE + Total TFLOPS from FLOPs/step when not reported directly
- Support --ane (train_large_ane) and --no-ane-extras flags
- Dynamic pipeline timing breakdown + CKPT_PATH per mode
Dynamic weight pipeline that eliminates the ~3.7s recompile-every-10-steps
bottleneck. Weights are passed via IOSurface spatial dimension instead of
baked as constants, so kernels compile once at startup (345ms) and run
indefinitely without exec() restart.
Key components:
- training_dynamic/ — full pipeline (config, IO, MIL generators, train loop)
- 9 dynamic kernels shared across all 12 layers
- Vocab compaction 32K→9.2K for faster classifier
- Vectorized cross-entropy with vDSP/NEON
- Adam optimizer with gradient clipping + cosine LR schedule
- Checkpoint save/resume
- test_dynamic_matmul.m — validates dynamic weight matmul vs cblas
- test_weight_patch.m — tests weight update via IOSurface
- dashboard.py — updated with --dynamic flag for v2 pipeline support,
improved step regex parsing, --scratch/--lr/--accum CLI args
Performance: 110ms/step steady-state (no recompile overhead)
ane_fwd=21 ane_bwd=28 io_fwd=12 io_bwd=15 silu=10 cls=13 rms=5 ms
Weave in scope notice near the top covering project intent, what it
is/isn't, hype clarification, maintenance expectations, and fork
encouragement. Consolidate private API disclaimer with existing
disclaimer section to avoid duplication.
https://claude.ai/code/session_01NNL4MVEY1aKp19eGHTYJUv
Key findings from running all 4 probes on Apple M5:
- Weight reload (unload+load after file overwrite) does NOT work — weights
are baked at compile time, output is identical regardless of file changes
- weightsBuffer IOSurface parameter also does not override compiled weights
- All QoS values 0-63 work, no measurable latency difference (~0.07ms/eval)
- _ANEPerformanceStats has hwExecutionTime (ns) + perfCounterData
- _ANEChainingRequest supports loopback execution (output→input chaining)
- _ANEClient has real-time eval path and chaining preparation methods
- procedureIndex 0-15 all succeed on single-procedure models
Fixed probe tests to use fp32 I/O with cast (matching inmem_peak pattern)
and 64+ channel kernels (ANE minimum size requirement).
Full analysis in training/m5result.md.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Four standalone probe tests to characterize the M5 ANE:
- test_weight_reload: Can weights be hot-swapped via unload+load without recompilation?
- test_perf_stats: Enumerate _ANEPerformanceStats methods/properties and hardware counters
- test_qos_sweep: Measure compile/load/eval latency across QoS 0-63
- test_ane_advanced: Probe SharedEvents, weightsBuffer IOSurface, procedureIndex, VirtualClient
Training telemetry (train_large.m):
- JSON lines to stderr with per-step timing breakdown and per-batch TFLOPS metrics
- Enables external monitoring tools to visualize ANE utilization in real-time
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Alvaro Videla
maderix
Manjeet Singh
04cb
Alvaro GPT
nabbilkhan
John Stephen Kromer
Guitared
Vipul
tastyheadphones
Claude
m0at
maderix