Add M1/M2 backward-compatibility module for ANE training

[codegen-sh]

Summary

Complete backward-compatibility module enabling ANE-accelerated training on M1/M2 hardware. Targets stability for 24/7 swarm use with zero regression on M4 paths.

Total: 1,180 lines of new code across 4 new files + Makefile update.

---

New Files

ane_hw_detect.h — Runtime chip detection (265 lines)

• ANEVersionDetect() reads sysctl hw.cpufamily, brand_string, hw.model — no private APIs, won't crash
• Chip profiles for M1/M2/M3/M4 with: supports_matmul, supports_sdpa, MIL version/target, max_compiles, IOSurface alignment, dimension limits
• Thread-safe singleton; falls back to conservative M2 defaults on unknown hardware
• Known cpufamily hashes for Firestorm (M1), Avalanche (M2), Everest (M3), and M4

ane_compat.h — M2-compatible MIL generators (407 lines)

• Every matmul → 1×1 conv: Linear projections become conv(weight=[out, in, 1, 1]) on tensor<fp16, [1, C, 1, S]> inputs
• program(1.0) with func main<ios16> target (vs M4's program(1.3) / ios18)
• Complete conv-only generators for:
  • Basic conv (mil_gen_conv_m2)
  • Fused QKV projections (3 parallel convs)
  • FFN up projections (W1 + W3 parallel convs)
  • SDPA forward with taps (conv projections + CPU attention fallback)
  • FFN forward with taps (RMSNorm + conv projections)
  • RMSNorm backward
  • Classifier forward/backward (transposed-weight conv for backward)
  • Final RMSNorm, softmax
• make_surface_m2() — 256-byte aligned IOSurface allocation
• Chip-aware dispatch: mil_gen_conv_compat() etc. return M2 variant or nil (caller uses M4 path)

ane_mem_budget.h — Conservative memory planner (211 lines)

• M2MemoryBudget(availableUnifiedGB=24) entry point
• Hard caps on M1/M2: batch=1, seq≤512, hidden≤4096
• Automatic gradient checkpointing with interval selection (2, 3, 4, or 6 layers)
• Falls back to seq_len reduction + fp16 grad accumulators if memory is still tight
• ANEAutoBudget() auto-detects available memory via hw.memsize

test_m2_compatibility.m — Stability test harness (297 lines)

• Full Stories110M 12-layer training loop
• 6-phase execution: HW detect → memory budget → model load → ANE compile → 30-min training → verdict
• Monitors: NaN/Inf with auto-recovery (LR halving), memory RSS, gradient norms, compile limit tracking
• Reports: ANE utilization %, estimated power draw, throughput (tok/s), step timing
• Pass criteria: 90% uptime, ≤2 NaN events, ≤3 eval failures, loss < 10.0

---

Modified Files

Makefile

• UNIVERSAL_CFLAGS with -arch arm64 (universal M1–M4)
• New HEADERS_COMPAT dependency group
• test_m2_compatibility target
• make all and make compat-check convenience targets

---

Architecture

ANEVersionDetect()       ← sysctl-based, runs once
       │
       ▼
ANEChipProfile.gen       ← M1/M2/M3/M4 enum
       │
   ┌───┴────┐
   │ M1/M2  │ M3/M4
   ▼        ▼
ane_compat.h    ane_mil_gen.h  (existing, untouched)
conv-only MIL   matmul MIL
program(1.0)    program(1.3)
ios16           ios18

M4 code paths are completely untouched. Dispatch happens via ane_is_m1_or_m2() checks in calling code; the existing M4 generators remain the default.

---

Estimated Performance (24 GB M2 MacBook)

Metric	M2 (conv-only)	M4 (matmul)	Ratio
Step time	~260 ms	~110 ms	~2.4× slower
Throughput	~100-150 tok/s	~250+ tok/s	~2.4×
Power (ANE)	~8W	~10W	0.8×
Peak memory	~8-10 GB	~6-8 GB	~1.3×
Viable for 24/7?	✅ Yes	✅ Yes	—

The 2.4× slowdown is expected: 1×1 conv on M2 ANE saturates fewer NE cores per op than M4's fused matmul, and attention core ops (Q·K^T, attn·V) fall back to CPU.

---

Risk Matrix

Risk	macOS 15.3	macOS 15.4	Mitigation
_ANEInMemoryModel API change	⚠️ Low	⚠️ Medium	Graceful fallback to CPU on compile failure
MIL program(1.0) deprecated	✅ Safe	⚠️ Low	Core ML still ships iOS 16 target support
Compile leak (M2 80 limit)	⚠️ Medium	⚠️ Medium	Tracked in test harness; exec() restart path
IOSurface alignment change	⚠️ Low	⚠️ Low	Runtime probing via chip profile
Conv op behavior change	✅ Safe	✅ Safe	Conv has been stable since ANE v1
Silent numerical drift (fp16)	⚠️ Medium	⚠️ Medium	NaN/Inf checks + grad norm monitoring
Memory pressure under swarm	⚠️ Medium	⚠️ Medium	Conservative budget + gradient checkpointing

---

Caveats

⚠️ Not validated on actual M1/M2 hardware — code is architecturally sound and structurally safe, but MIL syntax correctness and numerical stability need validation on real silicon.

Recommended next step: Run make test_m2_compatibility && ./test_m2_compatibility model.bin --duration=30 on an M2 MacBook and share the verdict output.

---

💻 View my work • 👤 Initiated by @dermitchell1993 • About Codegen
⛔ Remove Codegen from PR • 🚫 Ban action checks