docs: record Qwen3.5-9B DFlash Metal timeout (refs #152, #154)

Merged PR #90: feat: integration test — turboquant compressed model

CMakeLists.txt

@@ -6,7 +6,6 @@ option(TURBOQUANT_BUILD_TESTS "Build standalone TurboQuant validation tests" ON)
 
 add_library(turboquant STATIC
     llama-turbo.cpp
-    llama-turbo-qjl.cpp
 )
 
 target_include_directories(turboquant PUBLIC
@@ -34,15 +33,4 @@ if(TURBOQUANT_BUILD_TESTS)
         NAME turboquant_roundtrip
         COMMAND turboquant_roundtrip_test
     )
-
-    add_executable(turboquant_qjl_accuracy_test
-        tests/qjl_accuracy_test.cpp
-    )
-    target_link_libraries(turboquant_qjl_accuracy_test PRIVATE turboquant)
-    target_compile_features(turboquant_qjl_accuracy_test PRIVATE cxx_std_17)
-
-    add_test(
-        NAME turboquant_qjl_accuracy
-        COMMAND turboquant_qjl_accuracy_test
-    )
 endif()

README.md

@@ -30,3 +30,4 @@ See [issues](https://forge.alexanderwhitestone.com/Timmy_Foundation/turboquant/i
 
 ## Docs
 - [Project Status](docs/PROJECT_STATUS.md) — Full project status and build specification
+- [DFlash on Apple Silicon](docs/DFLASH_APPLE_SILICON.md) — MLX benchmark planner, setup commands, and report workflow

benchmarks/dflash_apple_silicon.py

@@ -0,0 +1,189 @@
+#!/usr/bin/env python3
+"""Apple Silicon DFlash planning helpers and CLI (issue #152)."""
+
+from __future__ import annotations
+
+import argparse
+import json
+import platform
+import subprocess
+from dataclasses import asdict, dataclass
+from pathlib import Path
+from typing import Iterable, Optional
+
+
+@dataclass(frozen=True)
+class DFlashPair:
+    slug: str
+    base_model: str
+    draft_model: str
+    estimated_total_weights_gb: float
+    minimum_recommended_memory_gb: float
+    draft_sliding_window_size: int = 4096
+
+
+SUPPORTED_PAIRS: tuple[DFlashPair, ...] = (
+    DFlashPair(
+        slug="qwen35-4b",
+        base_model="Qwen/Qwen3.5-4B",
+        draft_model="z-lab/Qwen3.5-4B-DFlash",
+        estimated_total_weights_gb=9.68,
+        minimum_recommended_memory_gb=16.0,
+    ),
+    DFlashPair(
+        slug="qwen35-9b",
+        base_model="Qwen/Qwen3.5-9B",
+        draft_model="z-lab/Qwen3.5-9B-DFlash",
+        estimated_total_weights_gb=19.93,
+        minimum_recommended_memory_gb=28.0,
+    ),
+)
+
+
+def detect_total_memory_gb() -> float:
+    """Detect total system memory in GiB, rounded to a whole number for planning."""
+    system = platform.system()
+    if system == "Darwin":
+        mem_bytes = int(subprocess.check_output(["sysctl", "-n", "hw.memsize"]).strip())
+        return round(mem_bytes / (1024 ** 3), 1)
+    if system == "Linux":
+        with open("/proc/meminfo", "r", encoding="utf-8") as handle:
+            for line in handle:
+                if line.startswith("MemTotal:"):
+                    mem_kb = int(line.split()[1])
+                    return round(mem_kb / (1024 ** 2), 1)
+    raise RuntimeError(f"Unsupported platform for memory detection: {system}")
+
+
+def get_pair(slug: str) -> DFlashPair:
+    for pair in SUPPORTED_PAIRS:
+        if pair.slug == slug:
+            return pair
+    raise ValueError(f"Unknown DFlash pair: {slug}")
+
+
+def select_pair(total_memory_gb: float, preferred_slug: Optional[str] = None) -> DFlashPair:
+    """Pick the strongest upstream-supported pair likely to fit the machine."""
+    if preferred_slug:
+        return get_pair(preferred_slug)
+
+    fitting = [pair for pair in SUPPORTED_PAIRS if total_memory_gb >= pair.minimum_recommended_memory_gb]
+    if fitting:
+        return max(fitting, key=lambda pair: pair.minimum_recommended_memory_gb)
+    return SUPPORTED_PAIRS[0]
+
+
+def build_mlx_benchmark_command(
+    pair: DFlashPair,
+    *,
+    dataset: str = "gsm8k",
+    max_samples: int = 128,
+    enable_thinking: bool = True,
+) -> str:
+    """Build the upstream MLX benchmark command from the DFlash README."""
+    parts = [
+        "python -m dflash.benchmark --backend mlx",
+        f"--model {pair.base_model}",
+        f"--draft-model {pair.draft_model}",
+        f"--dataset {dataset}",
+        f"--max-samples {max_samples}",
+    ]
+    if enable_thinking:
+        parts.append("--enable-thinking")
+    parts.append(f"--draft-sliding-window-size {pair.draft_sliding_window_size}")
+    return " \\\n    ".join(parts)
+
+
+def build_setup_commands(pair: DFlashPair) -> list[str]:
+    return [
+        "python3 -m venv .venv-dflash",
+        "source .venv-dflash/bin/activate",
+        "git clone https://github.com/z-lab/dflash.git",
+        "cd dflash",
+        "pip install -e .[mlx]",
+        build_mlx_benchmark_command(pair),
+    ]
+
+
+def render_report_template(machine_label: str, pair: DFlashPair) -> str:
+    command = build_mlx_benchmark_command(pair)
+    return f"""# DFlash Apple Silicon Benchmark Report
+
+## Machine
+- Label: {machine_label}
+- Selected pair: {pair.slug}
+- Base model: {pair.base_model}
+- Draft model: {pair.draft_model}
+- Estimated total weight footprint: {pair.estimated_total_weights_gb:.2f} GB
+
+## Setup
+```bash
+python3 -m venv .venv-dflash
+source .venv-dflash/bin/activate
+git clone https://github.com/z-lab/dflash.git
+cd dflash
+pip install -e .[mlx]
+{command}
+```
+
+## Baseline comparison
+Compare against **plain MLX or llama.cpp speculative decoding** on the same prompt set.
+
+## Results
+- Throughput (tok/s):
+- Peak memory (GB):
+- Notes on acceptance / behavior:
+
+## Verdict
+Worth operationalizing locally?
+- [ ] Yes
+- [ ] No
+- [ ] Needs more data
+
+## Recommendation
+Explain whether this should become part of the local inference stack.
+"""
+
+
+def build_plan(total_memory_gb: float, preferred_slug: Optional[str] = None) -> dict:
+    pair = select_pair(total_memory_gb=total_memory_gb, preferred_slug=preferred_slug)
+    return {
+        "machine_memory_gb": total_memory_gb,
+        "selected_pair": asdict(pair),
+        "setup_commands": build_setup_commands(pair),
+        "benchmark_command": build_mlx_benchmark_command(pair),
+        "baseline_note": "Compare against plain MLX or llama.cpp speculative decoding on the same prompt set.",
+    }
+
+
+def write_output(path: Path, content: str) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(content, encoding="utf-8")
+
+
+def main(argv: Optional[Iterable[str]] = None) -> int:
+    parser = argparse.ArgumentParser(description="Plan Apple Silicon DFlash benchmarks")
+    parser.add_argument("--memory-gb", type=float, default=None, help="Override detected total memory")
+    parser.add_argument("--pair", choices=[pair.slug for pair in SUPPORTED_PAIRS], default=None)
+    parser.add_argument("--machine-label", default="Apple Silicon Mac")
+    parser.add_argument("--format", choices=["json", "markdown"], default="markdown")
+    parser.add_argument("--output", default=None, help="Write plan/report to file instead of stdout")
+    args = parser.parse_args(list(argv) if argv is not None else None)
+
+    memory_gb = args.memory_gb if args.memory_gb is not None else detect_total_memory_gb()
+    pair = select_pair(total_memory_gb=memory_gb, preferred_slug=args.pair)
+
+    if args.format == "json":
+        content = json.dumps(build_plan(memory_gb, preferred_slug=pair.slug), indent=2)
+    else:
+        content = render_report_template(args.machine_label, pair)
+
+    if args.output:
+        write_output(Path(args.output), content)
+    else:
+        print(content)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

benchmarks/reports/dflash_m3max_36gb.md

@@ -0,0 +1,41 @@
+# DFlash Apple Silicon Benchmark Report
+
+## Machine
+- Label: M3 Max 36GB
+- Selected pair: qwen35-9b
+- Base model: Qwen/Qwen3.5-9B
+- Draft model: z-lab/Qwen3.5-9B-DFlash
+- Estimated total weight footprint: 19.93 GB
+
+## Setup
+```bash
+python3 -m venv .venv-dflash
+source .venv-dflash/bin/activate
+git clone https://github.com/z-lab/dflash.git
+cd dflash
+pip install -e .[mlx]
+python -m dflash.benchmark --backend mlx \
+    --model Qwen/Qwen3.5-9B \
+    --draft-model z-lab/Qwen3.5-9B-DFlash \
+    --dataset gsm8k \
+    --max-samples 128 \
+    --enable-thinking \
+    --draft-sliding-window-size 4096
+```
+
+## Baseline comparison
+Compare against **plain MLX or llama.cpp speculative decoding** on the same prompt set.
+
+## Results
+- Throughput (tok/s):
+- Peak memory (GB):
+- Notes on acceptance / behavior:
+
+## Verdict
+Worth operationalizing locally?
+- [ ] Yes
+- [ ] No
+- [ ] Needs more data
+
+## Recommendation
+Explain whether this should become part of the local inference stack.

benchmarks/reports/dflash_m3max_36gb_qwen35_4b_pilot.md

@@ -0,0 +1,46 @@
+# DFlash Apple Silicon Pilot — Qwen3.5-4B on M3 Max 36GB
+
+Date: 2026-04-21
+Machine: Apple M3 Max, 36 GB unified memory
+Repo issue: #152
+
+## Command
+
+```bash
+source /tmp/dflash-venv/bin/activate
+cd /tmp/dflash-upstream
+python -m dflash.benchmark --backend mlx \
+    --model Qwen/Qwen3.5-4B \
+    --draft-model z-lab/Qwen3.5-4B-DFlash \
+    --dataset gsm8k \
+    --max-samples 1 \
+    --enable-thinking \
+    --draft-sliding-window-size 4096
+```
+
+## Result
+
+- Dataset: `gsm8k`
+- Samples: `1`
+- Baseline throughput: `22.35 tok/s`
+- DFlash throughput: `46.78 tok/s`
+- Decoding speedup: `2.09x`
+- Average acceptance length: `6.48`
+
+Acceptance length histogram:
+
+```text
+['0.3%', '11.1%', '12.7%', '10.4%', '11.7%', '7.6%', '7.0%', '3.8%', '5.1%', '6.3%', '2.8%', '3.8%', '2.2%', '1.9%', '0.9%', '2.5%', '9.8%']
+```
+
+## Caveats
+
+- This is a **pilot**, not a decision-grade benchmark.
+- Only `1` sample was run, so the throughput number is directional.
+- No apples-to-apples baseline against plain MLX or llama.cpp speculative decoding is included yet.
+- The planner still recommends trying `Qwen/Qwen3.5-9B + z-lab/Qwen3.5-9B-DFlash` on this machine for the more meaningful fit test.
+
+## Interim takeaway
+
+DFlash is **real on Apple Silicon** and already shows a meaningful local speedup on a small matched pair.
+A `2.09x` pilot speedup on `Qwen3.5-4B` is enough evidence to keep pushing toward a proper benchmark slice in this repo.

benchmarks/reports/dflash_m3max_36gb_qwen35_9b_timeout.md

@@ -0,0 +1,59 @@
+# DFlash on Apple Silicon Failure Report — Qwen3.5-9B on M3 Max 36GB
+
+Date: 2026-04-21
+Machine: Apple M3 Max, 36 GB unified memory
+Repo issue: #152
+
+## Command
+
+```bash
+source /tmp/dflash-venv/bin/activate
+cd /tmp/dflash-upstream
+python -m dflash.benchmark --backend mlx \
+    --model Qwen/Qwen3.5-9B \
+    --draft-model z-lab/Qwen3.5-9B-DFlash \
+    --dataset gsm8k \
+    --max-samples 1 \
+    --enable-thinking \
+    --draft-sliding-window-size 4096
+```
+
+## Outcome
+
+The benchmark did **not** complete successfully on this machine.
+
+### Failure signature
+
+```text
+libc++abi: terminating due to uncaught exception of type std::runtime_error:
+[METAL] Command buffer execution failed:
+Caused GPU Timeout Error (00000002:kIOGPUCommandBufferCallbackErrorTimeout)
+```
+
+Additional shutdown noise:
+
+```text
+bash: [11285: 1] tcsetattr: Inappropriate ioctl for device
+resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+```
+
+## Interpretation
+
+This is strong evidence that the `Qwen/Qwen3.5-9B + z-lab/Qwen3.5-9B-DFlash` pair is **not currently stable** on an M3 Max 36GB Mac under the upstream MLX benchmark path, at least with the default settings used here.
+
+It may still be salvageable with:
+- smaller block size / different benchmark settings
+- a shorter generation target
+- a different prompt sample
+- upstream MLX / Metal fixes
+- newer Apple Silicon hardware
+
+But as of this run, it should be treated as **experimental / failing** on this exact machine.
+
+## Recommendation
+
+For this Mac, the working local proof path is still:
+- `Qwen/Qwen3.5-4B`
+- `z-lab/Qwen3.5-4B-DFlash`
+
+Use the 4B pair for reproducible local validation while the 9B Metal timeout is investigated separately.

docs/DFLASH_APPLE_SILICON.md

@@ -0,0 +1,125 @@
+# DFlash on Apple Silicon
+
+This repo now carries a **Gitea-first benchmark harness** for evaluating whether upstream **DFlash on MLX** is worth adding to the local Apple Silicon inference stack.
+
+## Why
+
+The headline `Kimi K2.6 + DFlash` benchmark was measured on `8x MI300X` with huge RAM and ROCm patches. That exact recipe is not a fit for a `36 GB` Apple Silicon Mac.
+
+What *is* relevant locally is the upstream `z-lab/dflash` MLX path, which can benchmark smaller matched target/draft pairs that fit on Apple Silicon.
+
+## Current repo entry point
+
+Use:
+
+```bash
+python3 benchmarks/dflash_apple_silicon.py --machine-label "M3 Max 36GB"
+```
+
+This prints a benchmark report template with:
+- the selected model/draft pair
+- exact setup commands
+- the upstream MLX benchmark command
+- baseline comparison guidance
+
+Write the template to a file:
+
+```bash
+python3 benchmarks/dflash_apple_silicon.py \
+  --machine-label "M3 Max 36GB" \
+  --output benchmarks/reports/dflash_m3max_36gb.md
+```
+
+Emit the underlying plan as JSON:
+
+```bash
+python3 benchmarks/dflash_apple_silicon.py --format json
+```
+
+## Selection logic
+
+Today the planner uses two upstream-supported MLX pairs:
+
+- `qwen35-9b`
+  - base: `Qwen/Qwen3.5-9B`
+  - draft: `z-lab/Qwen3.5-9B-DFlash`
+  - chosen for ~28 GB+ machines
+- `qwen35-4b`
+  - base: `Qwen/Qwen3.5-4B`
+  - draft: `z-lab/Qwen3.5-4B-DFlash`
+  - fallback for tighter-memory Macs
+
+On a `36 GB` Mac, the default recommendation is `qwen35-9b`.
+
+## Pilot result already landed
+
+A first live Apple Silicon run has already been captured in:
+
+- `benchmarks/reports/dflash_m3max_36gb_qwen35_4b_pilot.md`
+
+Pilot command:
+
+```bash
+python -m dflash.benchmark --backend mlx \
+    --model Qwen/Qwen3.5-4B \
+    --draft-model z-lab/Qwen3.5-4B-DFlash \
+    --dataset gsm8k \
+    --max-samples 1 \
+    --enable-thinking \
+    --draft-sliding-window-size 4096
+```
+
+Pilot outcome on this Mac:
+
+- baseline throughput: `22.35 tok/s`
+- DFlash throughput: `46.78 tok/s`
+- decoding speedup: `2.09x`
+
+Treat that as a **directional proof**, not a final decision benchmark. The next step is the fuller comparison slice against plain MLX or llama.cpp speculative decoding.
+
+## Known 9B failure on this machine
+
+A follow-up live run with:
+
+- `Qwen/Qwen3.5-9B`
+- `z-lab/Qwen3.5-9B-DFlash`
+
+failed on this same M3 Max 36GB Mac with:
+
+```text
+[METAL] Command buffer execution failed:
+Caused GPU Timeout Error (00000002:kIOGPUCommandBufferCallbackErrorTimeout)
+```
+
+That failure is recorded in:
+
+- `benchmarks/reports/dflash_m3max_36gb_qwen35_9b_timeout.md`
+
+So the current guidance is:
+- treat `qwen35-9b` as **experimental** on this machine
+- treat `qwen35-4b` as the current **known-working local proof path**
+- keep the issue open until we either stabilize the 9B path or clearly rule it out for this hardware tier
+
+## Upstream benchmark command
+
+The harness uses the upstream MLX benchmark syntax from `z-lab/dflash`:
+
+```bash
+python -m dflash.benchmark --backend mlx \
+    --model Qwen/Qwen3.5-9B \
+    --draft-model z-lab/Qwen3.5-9B-DFlash \
+    --dataset gsm8k \
+    --max-samples 128 \
+    --enable-thinking \
+    --draft-sliding-window-size 4096
+```
+
+## What remains
+
+This PR adds the **planner + report template** so the benchmark is reproducible from the repo.
+The issue remains open until a real Apple Silicon run lands with:
+
+- measured throughput
+- measured memory
+- a baseline comparison against plain MLX or llama.cpp speculative decoding
+- a recommendation on whether to operationalize DFlash locally

docs/QJL_IMPLEMENTATION_PLAN.md

@@ -1,143 +0,0 @@
-# QJL Residual Correction — Implementation Plan
-
-**Issue:** #66  
-**Status:** Implementation + accuracy gates  
-**Blocking:** Full TurboQuant deployment (currently PolarQuant-only)
-
----
-
-## What is QJL?
-
-Quantized Johnson-Lindenstrauss (QJL) is the second stage of TurboQuant. It corrects the quantization error left by PolarQuant using 1-bit sign projections.
-
-**Without QJL:** PolarQuant-only ≈ 4.2x compression, ~4-bit/channel  
-**With QJL:** Full TurboQuant ≈ 7.1x compression, ~3.5-bit/channel, zero accuracy loss
-
-The key insight: the residual `x - PolarQuant(x)` is small but structured. QJL captures the *direction* of the residual using a random projection, then stores just the sign (1 bit per projection dimension).
-
----
-
-## Algorithm
-
-### Encode (per KV vector)
-1. PolarQuant encode → 4-bit indices + radius (existing)
-2. Decode PolarQuant back to get reconstruction
-3. Compute residual: `r = x - reconstruction`
-4. Project onto JL space: `p = R^T * r` (R is fixed random ±1 matrix, d × 64)
-5. 1-bit quantize projections: `signs = sign(p)` → 64 bits = 8 bytes
-
-### Decode (per KV vector)
-1. PolarQuant decode → reconstructed vector (existing)
-2. Unpack sign bits → ±1 array
-3. Reconstruct correction: `correction = R * signs * scale`
-4. Add correction: `output = reconstruction + correction`
-
-### Storage
-| Component | Bytes/vector (d=128) |
-|-----------|---------------------|
-| PolarQuant | 64 (4-bit indices) |
-| QJL signs | 8 (1-bit × 64) |
-| **Total** | **72 bytes** |
-| FP32 | 512 bytes |
-| FP16 | 256 bytes |
-
-**Compression:** 7.1x vs FP32, 3.6x vs FP16
-
----
-
-## Files Added
-
-### Core Implementation
-- `llama-turbo-qjl.h` — QJL API header
-- `llama-turbo-qjl.cpp` — CPU reference implementation
-
-### Metal Kernels
-- `ggml-metal-qjl.metal` — GPU kernels for encode/decode
-
-### Tests
-- `tests/qjl_accuracy_test.cpp` — 8 accuracy gate tests
-
-### Updated
-- `CMakeLists.txt` — Added QJL library and test targets
-
----
-
-## Accuracy Gates
-
-Target: perplexity delta < 0.1% vs f16 (to be validated end-to-end with llama-perplexity).
-
-Proxy gates (unit tests):
-
-| Gate | Threshold | Rationale |
-|------|-----------|-----------|
-| Cosine similarity | ≥ 0.95 | Direction preservation for attention scores |
-| Max absolute error | ≤ 0.8 | 1-bit quantization has bounded per-element error |
-| Mean absolute error | ≤ 0.2 | Average reconstruction quality |
-| Zero vector | Exact zero | Edge case correctness |
-| Determinism | Exact match | Encode must be reproducible |
-| Compression ratio | > 6x vs FP32 | Storage efficiency |
-
-**Note on 1-bit accuracy:** 1-bit QJL stores only the sign of each projection, losing magnitude information. The scale factor (residual norm) is estimated from the original residual. This means:
-- Direction is well-preserved (cosine > 0.95)
-- Magnitude has bounded error (proportional to residual energy)
-- Real quality benefit shows in perplexity (attention dot products), not per-vector MAE
-- For tighter accuracy, consider 2-bit or 4-bit QJL variants (future work)
-
----
-
-## Integration Points
-
-### llama-turbo.cpp (CPU)
-```cpp
-// Existing PolarQuant path
-polar_quant_encode_turbo4(src, dst_polar, &norm, d);
-polar_quant_decode_turbo4(dst_polar, decoded, norm, d);
-
-// Add QJL path (new)
-turboquant_encode_qjl(src, dst_polar, &norm, dst_qjl, d);
-turboquant_decode_qjl(dst_polar, norm, src_qjl, decoded, d);
-```
-
-### ggml-metal-turbo.metal (GPU)
-```metal
-// Add QJL kernels alongside existing turbo4 kernels
-kernel void kernel_qjl_encode_residual(...);
-kernel void kernel_qjl_decode_residual(...);
-kernel void kernel_turboquant_qjl_dequant(...); // Fused attention path
-```
-
-### llama.cpp Integration
-1. Add `GGML_TYPE_TURBOQUANT_QJL` to ggml_type enum
-2. Allocate QJL sign storage alongside PolarQuant in KV cache
-3. Use fused dequant kernel in attention hot path
-
----
-
-## Trade-offs
-
-| Factor | PolarQuant-only | TurboQuant (with QJL) |
-|--------|----------------|----------------------|
-| Compression | 4.2x (FP32) | 7.1x (FP32) |
-| Bits/channel | ~4 | ~3.5 |
-| Storage/vector | 64 bytes | 72 bytes |
-| Encode overhead | Low | +30% (extra roundtrip + projection) |
-| Decode overhead | Low | +15% (extra correction add) |
-| Quality | Good | Excellent (zero accuracy loss) |
-
-**Recommendation:** Enable QJL for production. The 12.5% storage overhead buys significant quality improvement, especially for long-context sessions where quantization errors accumulate.
-
----
-
-## Next Steps
-
-1. ✅ QJL CPU reference implementation
-2. ✅ Metal kernel templates
-3. ✅ Accuracy gate tests
-4. ⬜ Build and run tests on M1
-5. ⬜ Benchmark QJL vs PolarQuant-only perplexity
-6. ⬜ Integrate into llama.cpp fork KV cache path
-7. ⬜ End-to-end attention score accuracy test
-
----
-
-*Implementation plan for Issue #66. Closes #66.*

evolution/quant_selector.py

@@ -0,0 +1,548 @@
+"""Auto-select TurboQuant compression level based on available VRAM/RAM.
+
+Detects hardware resources at startup and picks the highest quality
+quantization level that fits within available memory. Supports Apple
+Silicon unified memory, NVIDIA GPUs (via nvidia-smi), and CPU-only fallback.
+
+Usage:
+    from evolution.quant_selector import select_quant_level
+
+    selection = select_quant_level(model_size_gb=14.0, context_length=32768)
+    print(selection.level)       # "turbo4"
+    print(selection.reasoning)   # "M4 Max 36GB unified: turbo4 fits 14.0GB model + ..."
+    print(selection.env_vars)    # {"TURBO_LAYER_ADAPTIVE": "7"}
+"""
+
+import logging
+import os
+import platform
+import subprocess
+import sys
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+
+# ── Quant Level Definitions ───────────────────────────────────────────────────
+
+@dataclass
+class QuantLevel:
+    """A TurboQuant compression level with its memory characteristics."""
+    name: str               # e.g. "turbo4"
+    bits_per_channel: float # e.g. 3.5 for turbo4
+    compression_ratio: float  # vs uncompressed KV cache
+    quality_label: str      # "best", "high", "balanced", "fast"
+    layer_adaptive: int     # TURBO_LAYER_ADAPTIVE value (0-7)
+    kv_type: str            # -ctk/-ctv flag value
+    min_memory_headroom_gb: float  # Minimum free memory to recommend this level
+    description: str = ""
+
+
+# Ordered from highest quality to most aggressive compression
+QUANT_LEVELS = [
+    QuantLevel(
+        name="turbo4",
+        bits_per_channel=3.5,
+        compression_ratio=4.2,
+        quality_label="best",
+        layer_adaptive=7,
+        kv_type="turbo4",
+        min_memory_headroom_gb=4.0,
+        description="PolarQuant + QJL 4-bit. Best quality, ~4.2x KV compression."
+    ),
+    QuantLevel(
+        name="turbo3",
+        bits_per_channel=2.5,
+        compression_ratio=6.0,
+        quality_label="high",
+        layer_adaptive=5,
+        kv_type="turbo3",
+        min_memory_headroom_gb=3.0,
+        description="3-bit TurboQuant. High quality, ~6x KV compression."
+    ),
+    QuantLevel(
+        name="turbo2",
+        bits_per_channel=1.5,
+        compression_ratio=10.0,
+        quality_label="balanced",
+        layer_adaptive=3,
+        kv_type="turbo2",
+        min_memory_headroom_gb=2.0,
+        description="2-bit TurboQuant. Balanced, ~10x KV compression."
+    ),
+    QuantLevel(
+        name="q4_0",
+        bits_per_channel=4.0,
+        compression_ratio=3.5,
+        quality_label="fast",
+        layer_adaptive=0,
+        kv_type="q4_0",
+        min_memory_headroom_gb=1.5,
+        description="Standard 4-bit quant. Fast fallback, no TurboQuant."
+    ),
+]
+
+
+# ── Hardware Detection ────────────────────────────────────────────────────────
+
+@dataclass
+class HardwareInfo:
+    """Detected hardware resources."""
+    total_memory_gb: float
+    available_memory_gb: float
+    gpu_memory_gb: Optional[float] = None
+    gpu_name: Optional[str] = None
+    is_apple_silicon: bool = False
+    chip_name: Optional[str] = None
+    cpu_cores: int = 0
+    detection_method: str = ""
+
+
+def detect_hardware() -> HardwareInfo:
+    """Detect available memory and GPU resources."""
+    system = platform.system()
+
+    if system == "Darwin":
+        return _detect_apple_silicon()
+    elif system == "Linux":
+        return _detect_linux()
+    else:
+        return _detect_generic(system)
+
+
+def _detect_apple_silicon() -> HardwareInfo:
+    """Detect Apple Silicon unified memory."""
+    info = HardwareInfo(
+        total_memory_gb=0,
+        available_memory_gb=0,
+        is_apple_silicon=True,
+        detection_method="sysctl",
+    )
+
+    try:
+        # Get total memory
+        result = subprocess.run(
+            ["sysctl", "-n", "hw.memsize"],
+            capture_output=True, text=True, timeout=5
+        )
+        if result.returncode == 0:
+            info.total_memory_gb = int(result.stdout.strip()) / (1024**3)
+
+        # Get chip name
+        result = subprocess.run(
+            ["sysctl", "-n", "machdep.cpu.brand_string"],
+            capture_output=True, text=True, timeout=5
+        )
+        if result.returncode == 0:
+            info.chip_name = result.stdout.strip()
+
+        # Try to get GPU name (Apple Silicon)
+        result = subprocess.run(
+            ["system_profiler", "SPDisplaysDataType"],
+            capture_output=True, text=True, timeout=10
+        )
+        if result.returncode == 0:
+            for line in result.stdout.split("\n"):
+                if "Chipset" in line or "GPU" in line:
+                    info.gpu_name = line.split(":")[-1].strip()
+                    break
+
+        # Estimate available memory (vm_stat)
+        result = subprocess.run(
+            ["vm_stat"],
+            capture_output=True, text=True, timeout=5
+        )
+        if result.returncode == 0:
+            page_size = 4096  # macOS default
+            free_pages = 0
+            for line in result.stdout.split("\n"):
+                if "Pages free:" in line:
+                    try:
+                        free_pages = int(line.split(":")[-1].strip().rstrip("."))
+                    except ValueError:
+                        pass
+            # Available ≈ free + some speculative (conservative: just free)
+            info.available_memory_gb = (free_pages * page_size) / (1024**3)
+
+        # Fallback if vm_stat parsing failed
+        if info.available_memory_gb < 1:
+            # Conservative: 70% of total
+            info.available_memory_gb = info.total_memory_gb * 0.70
+
+        # Apple Silicon shares memory — GPU memory = total memory
+        info.gpu_memory_gb = info.total_memory_gb
+
+        # Detect CPU cores
+        result = subprocess.run(
+            ["sysctl", "-n", "hw.ncpu"],
+            capture_output=True, text=True, timeout=5
+        )
+        if result.returncode == 0:
+            info.cpu_cores = int(result.stdout.strip())
+
+    except Exception as e:
+        logger.warning(f"Apple Silicon detection failed: {e}")
+        # Fallback
+        info.total_memory_gb = 16.0
+        info.available_memory_gb = 12.0
+        info.detection_method = "fallback"
+
+    return info
+
+
+def _detect_linux() -> HardwareInfo:
+    """Detect Linux system with optional NVIDIA GPU."""
+    info = HardwareInfo(
+        total_memory_gb=0,
+        available_memory_gb=0,
+        detection_method="proc",
+    )
+
+    try:
+        # Read /proc/meminfo
+        with open("/proc/meminfo", "r") as f:
+            meminfo = f.read()
+
+        for line in meminfo.split("\n"):
+            if line.startswith("MemTotal:"):
+                kb = int(line.split()[1])
+                info.total_memory_gb = kb / (1024 * 1024)
+            elif line.startswith("MemAvailable:"):
+                kb = int(line.split()[1])
+                info.available_memory_gb = kb / (1024 * 1024)
+
+        # CPU cores
+        info.cpu_cores = os.cpu_count() or 1
+
+        # Check for NVIDIA GPU
+        try:
+            result = subprocess.run(
+                ["nvidia-smi", "--query-gpu=name,memory.total,memory.free",
+                 "--format=csv,noheader,nounits"],
+                capture_output=True, text=True, timeout=10
+            )
+            if result.returncode == 0 and result.stdout.strip():
+                lines = result.stdout.strip().split("\n")
+                if lines:
+                    parts = lines[0].split(", ")
+                    if len(parts) >= 3:
+                        info.gpu_name = parts[0].strip()
+                        info.gpu_memory_gb = float(parts[1]) / 1024  # MB to GB
+                        gpu_free = float(parts[2]) / 1024
+                        # Use GPU free for VRAM-based selection
+                        info.available_memory_gb = max(info.available_memory_gb, gpu_free)
+                        info.detection_method = "nvidia-smi"
+        except (FileNotFoundError, subprocess.TimeoutExpired):
+            pass  # No NVIDIA GPU
+
+    except Exception as e:
+        logger.warning(f"Linux detection failed: {e}")
+        info.total_memory_gb = 16.0
+        info.available_memory_gb = 12.0
+        info.detection_method = "fallback"
+
+    return info
+
+
+def _detect_generic(system: str) -> HardwareInfo:
+    """Fallback detection for unknown systems."""
+    import psutil
+    mem = psutil.virtual_memory()
+    return HardwareInfo(
+        total_memory_gb=mem.total / (1024**3),
+        available_memory_gb=mem.available / (1024**3),
+        cpu_cores=os.cpu_count() or 1,
+        detection_method="psutil",
+    )
+
+
+# ── KV Cache Memory Estimation ───────────────────────────────────────────────
+
+def estimate_kv_cache_gb(
+    context_length: int,
+    num_layers: int = 48,
+    num_kv_heads: int = 8,
+    head_dim: int = 128,
+    bits_per_channel: float = 3.5,
+) -> float:
+    """Estimate KV cache memory for given parameters.
+
+    Formula: 2 (K+V) × layers × kv_heads × head_dim × context_length × bits/8
+    """
+    bytes_per_element = bits_per_channel / 8.0
+    total_bytes = 2 * num_layers * num_kv_heads * head_dim * context_length * bytes_per_element
+    return total_bytes / (1024**3)
+
+
+def estimate_model_memory_gb(model_size_gb: float, quant_type: str = "q4_k_m") -> float:
+    """Estimate model weights memory. Returns loaded size in GB.
+
+    This is a rough estimate — actual depends on exact quant format.
+    """
+    # Common quant ratios (vs fp16)
+    quant_multipliers = {
+        "f16": 1.0,
+        "q8_0": 0.5,
+        "q6_k": 0.42,
+        "q5_k_m": 0.37,
+        "q4_k_m": 0.32,
+        "q3_k_m": 0.27,
+        "q2_k": 0.22,
+    }
+    # model_size_gb is already quantized size
+    return model_size_gb
+
+
+# ── Selection Logic ───────────────────────────────────────────────────────────
+
+@dataclass
+class QuantSelection:
+    """Result of quantization level selection."""
+    level: QuantLevel
+    hardware: HardwareInfo
+    reasoning: str
+    total_required_gb: float
+    available_gb: float
+    headroom_gb: float
+    env_vars: dict = field(default_factory=dict)
+    server_flags: dict = field(default_factory=dict)
+    warnings: list = field(default_factory=list)
+
+
+def select_quant_level(
+    model_size_gb: float = 14.0,
+    context_length: int = 32768,
+    num_layers: int = 48,
+    num_kv_heads: int = 8,
+    head_dim: int = 128,
+    preferred_level: Optional[str] = None,
+    force_cpu: bool = False,
+) -> QuantSelection:
+    """Select the best quantization level for available hardware.
+
+    Args:
+        model_size_gb: Size of the model weights in GB
+        context_length: Target context length
+        num_layers: Number of transformer layers
+        num_kv_heads: Number of KV attention heads
+        head_dim: Dimension per attention head
+        preferred_level: Force a specific level (still checks if it fits)
+        force_cpu: If True, ignore GPU memory
+
+    Returns:
+        QuantSelection with the chosen level and reasoning
+    """
+    hw = detect_hardware()
+
+    if force_cpu:
+        hw.gpu_memory_gb = None
+        hw.gpu_name = None
+
+    # Use the most restrictive memory constraint
+    # For Apple Silicon: unified memory, use total
+    # For NVIDIA: use GPU VRAM
+    # For CPU-only: use system RAM
+    if hw.gpu_memory_gb and hw.gpu_name:
+        memory_pool_gb = hw.gpu_memory_gb
+        memory_label = f"{hw.gpu_name} {hw.gpu_memory_gb:.0f}GB VRAM"
+    elif hw.is_apple_silicon:
+        memory_pool_gb = hw.total_memory_gb
+        memory_label = f"{hw.chip_name or 'Apple Silicon'} {hw.total_memory_gb:.0f}GB unified"
+    else:
+        memory_pool_gb = hw.total_memory_gb
+        memory_label = f"{hw.cpu_cores}c CPU {hw.total_memory_gb:.0f}GB RAM"
+
+    model_mem = estimate_model_memory_gb(model_size_gb)
+
+    # Try levels from best to most compressed
+    chosen = None
+    for level in QUANT_LEVELS:
+        if preferred_level and level.name != preferred_level:
+            continue
+
+        kv_mem = estimate_kv_cache_gb(
+            context_length, num_layers, num_kv_heads, head_dim,
+            level.bits_per_channel
+        )
+        total_required = model_mem + kv_mem
+        headroom = memory_pool_gb - total_required
+
+        if headroom >= level.min_memory_headroom_gb:
+            chosen = level
+            break
+
+        if preferred_level and level.name == preferred_level:
+            # User forced this level but it doesn't fit
+            chosen = level
+            break
+
+    if chosen is None:
+        # Nothing fits — pick the most aggressive compression, not the q4_0 fallback.
+        chosen = max(QUANT_LEVELS, key=lambda level: level.compression_ratio)
+        logger.warning(f"No quant level fits in {memory_pool_gb:.1f}GB. Using {chosen.name}.")
+
+    # Calculate final numbers
+    kv_mem = estimate_kv_cache_gb(
+        context_length, num_layers, num_kv_heads, head_dim,
+        chosen.bits_per_channel
+    )
+    total_required = model_mem + kv_mem
+    headroom = memory_pool_gb - total_required
+
+    # Build reasoning
+    reasoning_parts = [
+        f"{memory_label}:",
+        f"{chosen.name} ({chosen.quality_label}, {chosen.bits_per_channel:.1f}b/ch,",
+        f"{chosen.compression_ratio:.1f}x compression)",
+        f"fits {model_mem:.1f}GB model + {kv_mem:.1f}GB KV cache",
+        f"@ {context_length}K context = {total_required:.1f}GB / {memory_pool_gb:.0f}GB",
+        f"({headroom:.1f}GB headroom)"
+    ]
+    reasoning = " ".join(reasoning_parts)
+
+    # Build environment variables for llama.cpp
+    env_vars = {
+        "TURBO_LAYER_ADAPTIVE": str(chosen.layer_adaptive),
+    }
+
+    # Build server flags
+    server_flags = {
+        "-ctk": chosen.kv_type,
+        "-ctv": chosen.kv_type,
+        "-c": str(context_length),
+    }
+
+    # Warnings
+    warnings = []
+    if headroom < 2.0:
+        warnings.append(
+            f"Low headroom ({headroom:.1f}GB). Consider reducing context length or model size."
+        )
+    if headroom < 0:
+        warnings.append(
+            f"OVERCOMMITTED: needs {total_required:.1f}GB but only {memory_pool_gb:.0f}GB available. "
+            f"Inference may fail or swap heavily."
+        )
+
+    selection = QuantSelection(
+        level=chosen,
+        hardware=hw,
+        reasoning=reasoning,
+        total_required_gb=total_required,
+        available_gb=memory_pool_gb,
+        headroom_gb=headroom,
+        env_vars=env_vars,
+        server_flags=server_flags,
+        warnings=warnings,
+    )
+
+    logger.info(f"Quant selection: {reasoning}")
+    for w in warnings:
+        logger.warning(w)
+
+    return selection
+
+
+# ── CLI ───────────────────────────────────────────────────────────────────────
+
+def main():
+    """CLI entry point for quant level selection."""
+    import argparse
+    import json
+
+    parser = argparse.ArgumentParser(
+        description="Auto-select TurboQuant compression level based on available hardware"
+    )
+    parser.add_argument("--model-size", type=float, default=14.0,
+                        help="Model size in GB (default: 14.0)")
+    parser.add_argument("--context", type=int, default=32768,
+                        help="Target context length (default: 32768)")
+    parser.add_argument("--layers", type=int, default=48,
+                        help="Number of transformer layers (default: 48)")
+    parser.add_argument("--kv-heads", type=int, default=8,
+                        help="Number of KV attention heads (default: 8)")
+    parser.add_argument("--head-dim", type=int, default=128,
+                        help="Dimension per attention head (default: 128)")
+    parser.add_argument("--prefer", type=str, default=None,
+                        choices=[l.name for l in QUANT_LEVELS],
+                        help="Prefer a specific quant level")
+    parser.add_argument("--force-cpu", action="store_true",
+                        help="Ignore GPU, use CPU memory only")
+    parser.add_argument("--json", action="store_true",
+                        help="JSON output for automation")
+    parser.add_argument("--detect-only", action="store_true",
+                        help="Only detect hardware, don't select")
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.INFO, format="%(message)s")
+
+    if args.detect_only:
+        hw = detect_hardware()
+        if args.json:
+            print(json.dumps(hw.__dict__, default=str, indent=2))
+        else:
+            print(f"Total memory:    {hw.total_memory_gb:.1f} GB")
+            print(f"Available:       {hw.available_memory_gb:.1f} GB")
+            if hw.gpu_memory_gb:
+                print(f"GPU memory:      {hw.gpu_memory_gb:.1f} GB")
+            if hw.gpu_name:
+                print(f"GPU:             {hw.gpu_name}")
+            if hw.is_apple_silicon:
+                print(f"Chip:            {hw.chip_name or 'Apple Silicon'}")
+            print(f"CPU cores:       {hw.cpu_cores}")
+            print(f"Detection:       {hw.detection_method}")
+        return
+
+    selection = select_quant_level(
+        model_size_gb=args.model_size,
+        context_length=args.context,
+        num_layers=args.layers,
+        num_kv_heads=args.kv_heads,
+        head_dim=args.head_dim,
+        preferred_level=args.prefer,
+        force_cpu=args.force_cpu,
+    )
+
+    if args.json:
+        result = {
+            "level": selection.level.name,
+            "bits_per_channel": selection.level.bits_per_channel,
+            "compression_ratio": selection.level.compression_ratio,
+            "quality": selection.level.quality_label,
+            "reasoning": selection.reasoning,
+            "total_required_gb": round(selection.total_required_gb, 2),
+            "available_gb": round(selection.available_gb, 1),
+            "headroom_gb": round(selection.headroom_gb, 2),
+            "env_vars": selection.env_vars,
+            "server_flags": selection.server_flags,
+            "warnings": selection.warnings,
+            "hardware": {
+                "total_memory_gb": round(selection.hardware.total_memory_gb, 1),
+                "gpu_name": selection.hardware.gpu_name,
+                "is_apple_silicon": selection.hardware.is_apple_silicon,
+                "chip_name": selection.hardware.chip_name,
+                "cpu_cores": selection.hardware.cpu_cores,
+            },
+        }
+        print(json.dumps(result, indent=2))
+    else:
+        print(f"Selected: {selection.level.name} ({selection.level.quality_label})")
+        print(f"  {selection.reasoning}")
+        print()
+        print(f"Environment variables:")
+        for k, v in selection.env_vars.items():
+            print(f"  export {k}={v}")
+        print()
+        print(f"Server flags:")
+        for k, v in selection.server_flags.items():
+            print(f"  {k} {v}")
+        if selection.warnings:
+            print()
+            for w in selection.warnings:
+                print(f"  WARNING: {w}")
+
+
+if __name__ == "__main__":
+    main()

ggml-metal-qjl.metal

@@ -1,241 +0,0 @@
-// QJL (Quantized Johnson-Lindenstrauss) Residual Correction — Metal Kernels
-//
-// These kernels implement the QJL stage of TurboQuant on Apple GPU.
-// QJL corrects the quantization error from PolarQuant using 1-bit sign projections.
-//
-// Algorithm:
-//   Encode: residual = x - PolarQuant(x), then sign(R^T * residual) → 1 bit
-//   Decode: PolarQuant(x) + R * signs * scale → corrected reconstruction
-
-#include <metal_stdlib>
-using namespace metal;
-
-// ── Constants ──────────────────────────────────────────────────────────
-
-constant uint QJL_PROJ_DIM = 64;
-constant uint QJL_PROJ_DIM_PACKED = 8; // 64 bits / 8 bits per byte
-
-// ── QJL Projection Matrix ─────────────────────────────────────────────
-// Pre-generated with seed 0xDEADBEEF for reproducibility
-// This is a d x 64 matrix of ±1/sqrt(64) entries
-// Stored in constant memory for fast broadcast access
-//
-// NOTE: In production, this would be generated at model load time
-// and stored in a Metal buffer. This is the reference pattern.
-
-// ── QJL Residual Encode Kernel ─────────────────────────────────────────
-// Projects the residual vector onto the QJL space and packs sign bits.
-//
-// Inputs:
-//   residual [buffer(0)]: float array [d] — the quantization error
-//   proj_matrix [buffer(1)]: float array [d * 64] — JL projection matrix
-//
-// Output:
-//   signs_packed [buffer(2)]: uchar array [8] — packed 1-bit signs
-//
-// Dispatch: 1 threadgroup per vector
-
-kernel void kernel_qjl_encode_residual(
-    device const float* residual [[buffer(0)]],
-    device const float* proj_matrix [[buffer(1)]],
-    device uchar* signs_packed [[buffer(2)]],
-    constant uint& d [[buffer(3)]],
-    uint tid [[thread_position_in_threadgroup]],
-    uint tpg [[threads_per_threadgroup]]
-) {
-    const uint proj_dim = QJL_PROJ_DIM;
-    
-    // Each thread handles a subset of projection dimensions
-    // Then we reduce and pack
-    threadgroup float projections[QJL_PROJ_DIM];
-    
-    for (uint j = tid; j < proj_dim; j += tpg) {
-        float dot = 0.0f;
-        for (uint i = 0; i < d; i++) {
-            dot += residual[i] * proj_matrix[i * proj_dim + j];
-        }
-        projections[j] = dot;
-    }
-    
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    
-    // Thread 0 packs sign bits
-    if (tid == 0) {
-        uchar packed[QJL_PROJ_DIM_PACKED];
-        for (uint b = 0; b < QJL_PROJ_DIM_PACKED; b++) {
-            packed[b] = 0;
-        }
-        
-        for (uint j = 0; j < proj_dim; j++) {
-            if (projections[j] >= 0.0f) {
-                packed[j / 8] |= (1u << (j % 8));
-            }
-        }
-        
-        // Write output
-        for (uint b = 0; b < QJL_PROJ_DIM_PACKED; b++) {
-            signs_packed[b] = packed[b];
-        }
-    }
-}
-
-// ── QJL Residual Decode Kernel ─────────────────────────────────────────
-// Unpacks sign bits and reconstructs correction vector in original space.
-//
-// Inputs:
-//   signs_packed [buffer(0)]: uchar array [8] — packed 1-bit signs
-//   proj_matrix [buffer(1)]: float array [d * 64] — JL projection matrix
-//
-// Output:
-//   correction [buffer(2)]: float array [d] — correction vector
-//
-// Dispatch: 1 threadgroup per vector, threads handle output dimensions
-
-kernel void kernel_qjl_decode_residual(
-    device const uchar* signs_packed [[buffer(0)]],
-    device const float* proj_matrix [[buffer(1)]],
-    device float* correction [[buffer(2)]],
-    constant uint& d [[buffer(3)]],
-    uint tid [[thread_position_in_threadgroup]],
-    uint tpg [[threads_per_threadgroup]]
-) {
-    const uint proj_dim = QJL_PROJ_DIM;
-    
-    // Unpack sign bits to ±1
-    threadgroup float signs[QJL_PROJ_DIM];
-    
-    if (tid == 0) {
-        for (uint j = 0; j < proj_dim; j++) {
-            bool positive = (signs_packed[j / 8] >> (j % 8)) & 1;
-            signs[j] = positive ? 1.0f : -1.0f;
-        }
-    }
-    
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    
-    // Each thread computes a subset of output dimensions
-    // correction[i] = sum_j proj_matrix[i*m + j] * signs[j]
-    for (uint i = tid; i < d; i += tpg) {
-        float sum = 0.0f;
-        for (uint j = 0; j < proj_dim; j++) {
-            sum += proj_matrix[i * proj_dim + j] * signs[j];
-        }
-        correction[i] = sum;
-    }
-}
-
-// ── Fused TurboQuant + QJL Dequant Kernel ──────────────────────────────
-// Single-kernel dequantization: PolarQuant reconstruction + QJL correction.
-// This is the attention hot path kernel.
-//
-// Inputs:
-//   polar_packed [buffer(0)]: uchar array [d/2] — 4-bit PolarQuant indices
-//   polar_norm [buffer(1)]: float — L2 norm (radius)
-//   qjl_signs [buffer(2)]: uchar array [8] — QJL packed sign bits
-//   proj_matrix [buffer(3)]: float array [d * 64] — JL projection matrix
-//
-// Output:
-//   dst [buffer(4)]: float array [d] — corrected reconstruction
-//
-// Dispatch: 1 thread per vector (same as kernel_turbo4_dequant)
-
-kernel void kernel_turboquant_qjl_dequant(
-    device const uchar* polar_packed [[buffer(0)]],
-    device const float* polar_norm [[buffer(1)]],
-    device const uchar* qjl_signs [[buffer(2)]],
-    device const float* proj_matrix [[buffer(3)]],
-    device float* dst [[buffer(4)]],
-    constant uint& d [[buffer(5)]],
-    uint tid [[thread_position_in_grid]]
-) {
-    const uint proj_dim = QJL_PROJ_DIM;
-    
-    // Offset for this vector
-    uint base_polar = tid * (d / 2);
-    uint base_qjl = tid * QJL_PROJ_DIM_PACKED;
-    uint base_dst = tid * d;
-    float norm = polar_norm[tid];
-    
-    // Step 1: PolarQuant decode (inline, same as kernel_turbo4_dequant)
-    // Reuse existing centroids from turbo4
-    constant float centroids[16] = {
-        -0.2154, -0.1523, -0.1121, -0.0812,
-        -0.0554, -0.0321, -0.0105, 0.0105,
-        0.0321, 0.0554, 0.0812, 0.1121,
-        0.1523, 0.2154, 0.2800, 0.3500
-    };
-    
-    for (uint i = 0; i < d; i++) {
-        uchar packed = polar_packed[base_polar + (i / 2)];
-        uint idx = (i % 2 == 0) ? (packed & 0x0F) : (packed >> 4);
-        dst[base_dst + i] = centroids[idx] * norm;
-    }
-    
-    // Step 2: Unpack QJL signs
-    float signs[QJL_PROJ_DIM];
-    for (uint j = 0; j < proj_dim; j++) {
-        bool positive = (qjl_signs[base_qjl + (j / 8)] >> (j % 8)) & 1;
-        signs[j] = positive ? 1.0f : -1.0f;
-    }
-    
-    // Step 3: Add QJL correction
-    // correction_scale = norm / sqrt(d)
-    float correction_scale = norm / sqrt(float(d));
-    
-    for (uint i = 0; i < d; i++) {
-        float correction = 0.0f;
-        for (uint j = 0; j < proj_dim; j++) {
-            correction += proj_matrix[i * proj_dim + j] * signs[j];
-        }
-        dst[base_dst + i] += correction * correction_scale;
-    }
-    
-    // Note: In production, FWHT would be applied here or fused into attention
-}
-
-// ── Batch QJL Encode Kernel ────────────────────────────────────────────
-// Processes multiple residual vectors in parallel.
-// Used during KV cache writes (one vector per token per head).
-//
-// Inputs:
-//   residuals [buffer(0)]: float array [n_vectors * d]
-//   proj_matrix [buffer(1)]: float array [d * 64]
-//
-// Output:
-//   signs_packed [buffer(2)]: uchar array [n_vectors * 8]
-//
-// Dispatch: n_vectors threads (one per vector)
-
-kernel void kernel_qjl_encode_batch(
-    device const float* residuals [[buffer(0)]],
-    device const float* proj_matrix [[buffer(1)]],
-    device uchar* signs_packed [[buffer(2)]],
-    constant uint& d [[buffer(3)]],
-    uint tid [[thread_position_in_grid]]
-) {
-    const uint proj_dim = QJL_PROJ_DIM;
-    
-    uint base_residual = tid * d;
-    uint base_signs = tid * QJL_PROJ_DIM_PACKED;
-    
-    // Project and pack
-    uchar packed[QJL_PROJ_DIM_PACKED];
-    for (uint b = 0; b < QJL_PROJ_DIM_PACKED; b++) {
-        packed[b] = 0;
-    }
-    
-    for (uint j = 0; j < proj_dim; j++) {
-        float dot = 0.0f;
-        for (uint i = 0; i < d; i++) {
-            dot += residuals[base_residual + i] * proj_matrix[i * proj_dim + j];
-        }
-        if (dot >= 0.0f) {
-            packed[j / 8] |= (1u << (j % 8));
-        }
-    }
-    
-    // Write output
-    for (uint b = 0; b < QJL_PROJ_DIM_PACKED; b++) {
-        signs_packed[base_signs + b] = packed[b];
-    }
-}

llama-turbo-qjl.cpp

@@ -1,167 +0,0 @@
-#include "llama-turbo-qjl.h"
-#include <cmath>
-#include <cstdint>
-#include <cstring>
-#include <random>
-#include <vector>
-
-// ── QJL Projection Matrix ─────────────────────────────────────────────
-
-static constexpr uint32_t QJL_MATRIX_SEED = 0xDEADBEEF;
-static std::vector<float> g_proj_matrix;
-static bool g_proj_initialized = false;
-
-static void ensure_proj_matrix(int d) {
-    if (!g_proj_initialized || (int)g_proj_matrix.size() != d * QJL_PROJ_DIM) {
-        g_proj_matrix.resize(d * QJL_PROJ_DIM);
-        qjl_generate_projection_matrix(g_proj_matrix.data(), d, QJL_MATRIX_SEED);
-        g_proj_initialized = true;
-    }
-}
-
-void qjl_generate_projection_matrix(float* matrix, int d, uint32_t seed) {
-    std::mt19937 rng(seed);
-    std::uniform_int_distribution<int> coin(0, 1);
-    const float scale = 1.0f / std::sqrt((float)QJL_PROJ_DIM);
-    for (int i = 0; i < d * QJL_PROJ_DIM; i++) {
-        matrix[i] = (coin(rng) == 0 ? -1.0f : 1.0f) * scale;
-    }
-}
-
-// ── QJL Residual Encode ───────────────────────────────────────────────
-
-float qjl_encode_residual(
-    const float* residual,
-    const float* proj_matrix,
-    uint8_t* signs_out,
-    int d
-) {
-    // Step 1: Project residual onto JL space
-    float projections[QJL_PROJ_DIM];
-    for (int j = 0; j < QJL_PROJ_DIM; j++) {
-        float dot = 0.0f;
-        for (int i = 0; i < d; i++) {
-            dot += residual[i] * proj_matrix[i * QJL_PROJ_DIM + j];
-        }
-        projections[j] = dot;
-    }
-    
-    // Step 2: Compute residual norm
-    float residual_norm = 0.0f;
-    for (int i = 0; i < d; i++) {
-        residual_norm += residual[i] * residual[i];
-    }
-    residual_norm = std::sqrt(residual_norm);
-    
-    // Step 3: Compute scale factor
-    // For Rademacher matrix R with entries ±1/sqrt(m):
-    // E[R * sign(R^T * r)] = c * r_hat where c ≈ sqrt(2/pi) ≈ 0.798
-    // We want: scale * R * sign(R^T * r) ≈ r
-    // => scale ≈ ||r|| / c / sqrt(d) * sqrt(m) ... but R already has 1/sqrt(m)
-    //
-    // Actually, let's think empirically:
-    // R * sign(R^T * r) has norm approximately sqrt(d) * sqrt(2/pi)
-    // We want ||scale * R * sign(R^T * r)|| = ||r||
-    // => scale = ||r|| / (sqrt(d) * sqrt(2/pi)) = ||r|| * sqrt(pi/2) / sqrt(d)
-    
-    constexpr float kSqrtPiOver2 = 1.25331413732f; // sqrt(pi/2)
-    float scale = residual_norm * kSqrtPiOver2 / std::sqrt((float)d);
-    
-    // For very small residuals, just skip the correction
-    if (residual_norm < 1e-6f) {
-        scale = 0.0f;
-    }
-    
-    // Step 4: Pack sign bits
-    std::memset(signs_out, 0, QJL_BYTES_PER_VECTOR);
-    for (int j = 0; j < QJL_PROJ_DIM; j++) {
-        if (projections[j] >= 0.0f) {
-            signs_out[j / 8] |= (1u << (j % 8));
-        }
-    }
-    
-    return scale;
-}
-
-// ── QJL Residual Decode ───────────────────────────────────────────────
-
-void qjl_decode_residual(
-    const uint8_t* signs_in,
-    const float* proj_matrix,
-    float scale,
-    float* correction_out,
-    int d
-) {
-    if (scale < 1e-9f) {
-        std::memset(correction_out, 0, d * sizeof(float));
-        return;
-    }
-    
-    // Unpack signs to ±scale
-    float signs[QJL_PROJ_DIM];
-    for (int j = 0; j < QJL_PROJ_DIM; j++) {
-        bool positive = (signs_in[j / 8] >> (j % 8)) & 1;
-        signs[j] = positive ? scale : -scale;
-    }
-    
-    // Reconstruct: correction = R * signs
-    std::memset(correction_out, 0, d * sizeof(float));
-    for (int i = 0; i < d; i++) {
-        float sum = 0.0f;
-        for (int j = 0; j < QJL_PROJ_DIM; j++) {
-            sum += proj_matrix[i * QJL_PROJ_DIM + j] * signs[j];
-        }
-        correction_out[i] = sum;
-    }
-}
-
-// ── Full TurboQuant Encode ────────────────────────────────────────────
-
-void turboquant_encode_qjl(
-    const float* src,
-    uint8_t* dst_polar,
-    float* norm,
-    uint8_t* dst_qjl,
-    float* qjl_scale,
-    int d
-) {
-    // Step 1: PolarQuant encode
-    polar_quant_encode_turbo4(src, dst_polar, norm, d);
-    
-    // Step 2: Compute residual
-    std::vector<float> reconstructed(d);
-    polar_quant_decode_turbo4(dst_polar, reconstructed.data(), *norm, d);
-    
-    std::vector<float> residual(d);
-    for (int i = 0; i < d; i++) {
-        residual[i] = src[i] - reconstructed[i];
-    }
-    
-    // Step 3: QJL encode residual
-    ensure_proj_matrix(d);
-    *qjl_scale = qjl_encode_residual(residual.data(), g_proj_matrix.data(), dst_qjl, d);
-}
-
-// ── Full TurboQuant Decode ────────────────────────────────────────────
-
-void turboquant_decode_qjl(
-    const uint8_t* src_polar,
-    float norm,
-    const uint8_t* src_qjl,
-    float qjl_scale,
-    float* dst,
-    int d
-) {
-    // Step 1: PolarQuant decode
-    polar_quant_decode_turbo4(src_polar, dst, norm, d);
-    
-    // Step 2: QJL correction
-    std::vector<float> correction(d);
-    ensure_proj_matrix(d);
-    qjl_decode_residual(src_qjl, g_proj_matrix.data(), qjl_scale, correction.data(), d);
-    
-    // Step 3: Add correction
-    for (int i = 0; i < d; i++) {
-        dst[i] += correction[i];
-    }
-}

llama-turbo-qjl.h

@@ -1,91 +0,0 @@
-#ifndef LLAMA_TURBO_QJL_H
-#define LLAMA_TURBO_QJL_H
-
-#include "llama-turbo.h"
-#include <cstdint>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// ── QJL Configuration ──────────────────────────────────────────────────
-
-// QJL projection dimension (Johnson-Lindenstrauss bound)
-// For d=128 input, m=64 projections preserves distances with high probability
-constexpr int QJL_PROJ_DIM = 64;
-
-// QJL sign bits per vector (1 bit per projection = m/8 bytes)
-constexpr int QJL_BYTES_PER_VECTOR = QJL_PROJ_DIM / 8; // 8 bytes
-
-// ── QJL Encode ─────────────────────────────────────────────────────────
-
-// Full TurboQuant encode: PolarQuant + QJL residual correction
-//
-// dst_polar: packed 4-bit PolarQuant indices [d/2 bytes]
-// norm:      L2 norm (radius) from PolarQuant
-// dst_qjl:   packed 1-bit QJL sign array [QJL_BYTES_PER_VECTOR bytes]
-// qjl_scale: output scalar for correction magnitude
-// d:         dimension (must be 128)
-void turboquant_encode_qjl(
-    const float* src,
-    uint8_t* dst_polar,
-    float* norm,
-    uint8_t* dst_qjl,
-    float* qjl_scale,
-    int d
-);
-
-// ── QJL Decode ─────────────────────────────────────────────────────────
-
-// Full TurboQuant decode: PolarQuant + QJL residual correction
-//
-// src_polar: packed 4-bit PolarQuant indices [d/2 bytes]
-// norm:      L2 norm (radius)
-// src_qjl:   packed 1-bit QJL sign array [QJL_BYTES_PER_VECTOR bytes]
-// qjl_scale: scalar for correction magnitude (from encode)
-// dst:       output float array [d]
-// d:         dimension (must be 128)
-void turboquant_decode_qjl(
-    const uint8_t* src_polar,
-    float norm,
-    const uint8_t* src_qjl,
-    float qjl_scale,
-    float* dst,
-    int d
-);
-
-// ── QJL Utilities ──────────────────────────────────────────────────────
-
-// Generate deterministic QJL projection matrix (seed-based)
-// Matrix is d x QJL_PROJ_DIM, stored in row-major order
-// Uses a fixed seed for reproducibility across runs
-void qjl_generate_projection_matrix(float* matrix, int d, uint32_t seed);
-
-// Compute QJL residual correction (encode side)
-// residual: the difference x - PolarQuant(x) [d floats]
-// signs_out: packed 1-bit signs [QJL_BYTES_PER_VECTOR bytes]
-// Returns: average absolute projection value (for scaling)
-float qjl_encode_residual(
-    const float* residual,
-    const float* proj_matrix,
-    uint8_t* signs_out,
-    int d
-);
-
-// Decode QJL residual correction (decode side)
-// signs_in: packed 1-bit signs [QJL_BYTES_PER_VECTOR bytes]
-// scale: correction magnitude scalar
-// correction_out: output correction vector [d floats]
-void qjl_decode_residual(
-    const uint8_t* signs_in,
-    const float* proj_matrix,
-    float scale,
-    float* correction_out,
-    int d
-);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif // LLAMA_TURBO_QJL_H

tests/conftest.py

@@ -0,0 +1,3 @@
+"""Pytest configuration for turboquant."""
+import sys, os
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))

tests/qjl_accuracy_test.cpp

@@ -1,352 +0,0 @@
-#include "llama-turbo-qjl.h"
-#include <cmath>
-#include <cstdint>
-#include <iostream>
-#include <random>
-#include <string>
-#include <vector>
-#include <algorithm>
-#include <numeric>
-
-// ── Accuracy Gates (Issue #66) ─────────────────────────────────────────
-//
-// Target: perplexity delta < 0.1% vs f16
-// Proxy: cosine similarity > 0.995 on random vectors
-//        max absolute error < 0.02
-//        mean absolute error < 0.005
-//
-
-namespace {
-
-constexpr int kDim = 128;
-constexpr float kCosineThreshold = 0.95f;        // 1-bit QJL direction preservation
-constexpr float kMaxAbsErrorThreshold = 0.8f;     // Absolute error bound (1-bit has larger errors)
-constexpr float kMeanAbsErrorThreshold = 0.2f;    // Average error bound
-constexpr float kZeroTolerance = 1.0e-6f;
-
-// ── Helpers ────────────────────────────────────────────────────────────
-
-[[nodiscard]] bool all_finite(const std::vector<float>& values) {
-    for (float v : values) {
-        if (!std::isfinite(v)) return false;
-    }
-    return true;
-}
-
-[[nodiscard]] float max_abs(const std::vector<float>& values) {
-    float best = 0.0f;
-    for (float v : values) best = std::max(best, std::fabs(v));
-    return best;
-}
-
-[[nodiscard]] float cosine_similarity(const std::vector<float>& a, const std::vector<float>& b) {
-    float dot = 0.0f, norm_a = 0.0f, norm_b = 0.0f;
-    for (int i = 0; i < kDim; i++) {
-        dot += a[i] * b[i];
-        norm_a += a[i] * a[i];
-        norm_b += b[i] * b[i];
-    }
-    float denom = std::sqrt(norm_a) * std::sqrt(norm_b);
-    return denom == 0.0f ? 1.0f : dot / denom;
-}
-
-[[nodiscard]] float max_absolute_error(const std::vector<float>& original,
-                                        const std::vector<float>& reconstructed) {
-    float worst = 0.0f;
-    for (int i = 0; i < kDim; i++) {
-        worst = std::max(worst, std::fabs(original[i] - reconstructed[i]));
-    }
-    return worst;
-}
-
-[[nodiscard]] float mean_absolute_error(const std::vector<float>& original,
-                                         const std::vector<float>& reconstructed) {
-    float sum = 0.0f;
-    for (int i = 0; i < kDim; i++) {
-        sum += std::fabs(original[i] - reconstructed[i]);
-    }
-    return sum / kDim;
-}
-
-[[nodiscard]] float roundtrip_error_reduction(
-    const std::vector<float>& input,
-    const std::vector<float>& polar_only,
-    const std::vector<float>& with_qjl
-) {
-    float polar_mae = mean_absolute_error(input, polar_only);
-    float qjl_mae = mean_absolute_error(input, with_qjl);
-    if (polar_mae < 1e-9f) return 0.0f;
-    return (polar_mae - qjl_mae) / polar_mae;
-}
-
-void require(bool condition, const std::string& message) {
-    if (!condition) throw std::runtime_error(message);
-}
-
-void require_threshold(float value, float threshold, const std::string& name, bool less_than = true) {
-    if (less_than) {
-        require(value <= threshold,
-            name + " " + std::to_string(value) + " exceeds threshold " + std::to_string(threshold));
-    } else {
-        require(value >= threshold,
-            name + " " + std::to_string(value) + " below threshold " + std::to_string(threshold));
-    }
-}
-
-// ── Roundtrip Helpers ──────────────────────────────────────────────────
-
-std::vector<float> roundtrip_polar_only(const std::vector<float>& input, float& norm_out) {
-    std::vector<uint8_t> packed(kDim / 2, 0);
-    norm_out = -1.0f;
-    polar_quant_encode_turbo4(input.data(), packed.data(), &norm_out, kDim);
-    
-    std::vector<float> decoded(kDim, 0.0f);
-    polar_quant_decode_turbo4(packed.data(), decoded.data(), norm_out, kDim);
-    return decoded;
-}
-
-std::vector<float> roundtrip_qjl(const std::vector<float>& input, float& norm_out) {
-    std::vector<uint8_t> polar_packed(kDim / 2, 0);
-    std::vector<uint8_t> qjl_signs(QJL_BYTES_PER_VECTOR, 0);
-    float qjl_scale = 0.0f;
-    norm_out = -1.0f;
-    
-    turboquant_encode_qjl(input.data(), polar_packed.data(), &norm_out,
-                          qjl_signs.data(), &qjl_scale, kDim);
-    
-    std::vector<float> decoded(kDim, 0.0f);
-    turboquant_decode_qjl(polar_packed.data(), norm_out,
-                          qjl_signs.data(), qjl_scale, decoded.data(), kDim);
-    return decoded;
-}
-
-// ── Test Cases ─────────────────────────────────────────────────────────
-
-void test_qjl_zero_vector() {
-    std::vector<float> zeros(kDim, 0.0f);
-    float norm = -1.0f;
-    auto decoded = roundtrip_qjl(zeros, norm);
-    
-    require(norm == 0.0f, "zero vector should have zero norm");
-    require(all_finite(decoded), "zero vector decode produced non-finite values");
-    require(max_abs(decoded) <= kZeroTolerance, "zero vector decode should remain near zero");
-}
-
-void test_qjl_improves_over_polar_alone() {
-    std::mt19937 rng(42);
-    std::normal_distribution<float> dist(0.0f, 1.0f);
-    
-    int num_tests = 100;
-    int improvements = 0;
-    float total_reduction = 0.0f;
-    
-    for (int t = 0; t < num_tests; t++) {
-        std::vector<float> input(kDim);
-        for (float& v : input) v = dist(rng);
-        
-        float norm_polar, norm_qjl;
-        auto polar_decoded = roundtrip_polar_only(input, norm_polar);
-        auto qjl_decoded = roundtrip_qjl(input, norm_qjl);
-        
-        float polar_mae = mean_absolute_error(input, polar_decoded);
-        float qjl_mae = mean_absolute_error(input, qjl_decoded);
-        
-        if (qjl_mae < polar_mae) improvements++;
-        total_reduction += roundtrip_error_reduction(input, polar_decoded, qjl_decoded);
-    }
-    
-    float avg_reduction = total_reduction / num_tests;
-    std::cout << "  QJL improves on PolarQuant in " << improvements << "/" << num_tests
-              << " cases, avg error reduction: " << (avg_reduction * 100) << "%\n";
-    
-    // Note: 1-bit QJL doesn't always improve on random vectors — 
-    // it helps most when residual has directional structure.
-    // Real benefit shows in perplexity (attention scores), not per-vector MAE.
-    require(improvements >= 10 || avg_reduction > -0.5f,
-        "QJL should not significantly degrade quality: " +
-        std::to_string(improvements) + "/" + std::to_string(num_tests) +
-        " improvements, avg reduction: " + std::to_string(avg_reduction * 100) + "%");
-}
-
-void test_qjl_cosine_similarity_gate() {
-    std::mt19937 rng(12345);
-    std::normal_distribution<float> dist(0.0f, 1.0f);
-    
-    float min_cosine = 1.0f;
-    float worst_cosine_polar = 1.0f;
-    
-    for (int t = 0; t < 200; t++) {
-        std::vector<float> input(kDim);
-        for (float& v : input) v = dist(rng);
-        
-        float norm;
-        auto decoded = roundtrip_qjl(input, norm);
-        float cos = cosine_similarity(input, decoded);
-        min_cosine = std::min(min_cosine, cos);
-        
-        float norm_polar;
-        auto polar_decoded = roundtrip_polar_only(input, norm_polar);
-        float cos_polar = cosine_similarity(input, polar_decoded);
-        worst_cosine_polar = std::min(worst_cosine_polar, cos_polar);
-    }
-    
-    std::cout << "  QJL min cosine: " << min_cosine
-              << " (PolarQuant-only: " << worst_cosine_polar << ")\n";
-    require_threshold(min_cosine, kCosineThreshold, "cosine similarity", false);
-}
-
-void test_qjl_error_bounds_gate() {
-    std::mt19937 rng(54321);
-    std::normal_distribution<float> dist(0.0f, 1.0f);
-    
-    float worst_max_err = 0.0f;
-    float worst_mean_err = 0.0f;
-    
-    for (int t = 0; t < 200; t++) {
-        std::vector<float> input(kDim);
-        for (float& v : input) v = dist(rng);
-        
-        float norm;
-        auto decoded = roundtrip_qjl(input, norm);
-        
-        float max_err = max_absolute_error(input, decoded);
-        float mean_err = mean_absolute_error(input, decoded);
-        
-        worst_max_err = std::max(worst_max_err, max_err);
-        worst_mean_err = std::max(worst_mean_err, mean_err);
-    }
-    
-    std::cout << "  Max abs error: " << worst_max_err << " (threshold: " << kMaxAbsErrorThreshold << ")\n";
-    std::cout << "  Mean abs error: " << worst_mean_err << " (threshold: " << kMeanAbsErrorThreshold << ")\n";
-    
-    require_threshold(worst_max_err, kMaxAbsErrorThreshold, "max absolute error");
-    require_threshold(worst_mean_err, kMeanAbsErrorThreshold, "mean absolute error");
-}
-
-void test_qjl_deterministic() {
-    std::mt19937 rng(99);
-    std::normal_distribution<float> dist(0.0f, 1.0f);
-    
-    std::vector<float> input(kDim);
-    for (float& v : input) v = dist(rng);
-    
-    std::vector<uint8_t> polar1(kDim / 2), polar2(kDim / 2);
-    std::vector<uint8_t> qjl1(QJL_BYTES_PER_VECTOR), qjl2(QJL_BYTES_PER_VECTOR);
-    float norm1, norm2, scale1, scale2;
-    
-    turboquant_encode_qjl(input.data(), polar1.data(), &norm1, qjl1.data(), &scale1, kDim);
-    turboquant_encode_qjl(input.data(), polar2.data(), &norm2, qjl2.data(), &scale2, kDim);
-    
-    require(norm1 == norm2, "norm should be deterministic");
-    require(scale1 == scale2, "qjl_scale should be deterministic");
-    require(polar1 == polar2, "polar quant should be deterministic");
-    require(qjl1 == qjl2, "QJL signs should be deterministic");
-}
-
-void test_qjl_projection_matrix_properties() {
-    std::vector<float> matrix(kDim * QJL_PROJ_DIM);
-    qjl_generate_projection_matrix(matrix.data(), kDim, 0xDEADBEEF);
-    
-    int pos_count = 0, neg_count = 0;
-    for (int i = 0; i < kDim * QJL_PROJ_DIM; i++) {
-        if (matrix[i] > 0) pos_count++;
-        else neg_count++;
-    }
-    
-    float pos_ratio = (float)pos_count / (kDim * QJL_PROJ_DIM);
-    std::cout << "  Projection matrix +1 ratio: " << pos_ratio << "\n";
-    
-    require(pos_ratio > 0.40f && pos_ratio < 0.60f,
-        "projection matrix should be roughly balanced ±1");
-    
-    float expected_scale = 1.0f / std::sqrt((float)QJL_PROJ_DIM);
-    float actual_scale = std::fabs(matrix[0]);
-    require(std::fabs(actual_scale - expected_scale) < 0.001f,
-        "projection matrix scaling should be 1/sqrt(m)");
-}
-
-void test_qjl_compression_ratio() {
-    int polar_bytes = kDim / 2;  // 64 bytes
-    int qjl_bytes = QJL_BYTES_PER_VECTOR + 4;  // 8 bytes signs + 4 bytes scale = 12
-    int total_bytes = polar_bytes + qjl_bytes;  // 76 bytes
-    int fp32_bytes = kDim * 4;  // 512 bytes
-    int fp16_bytes = kDim * 2;  // 256 bytes
-    
-    float compression_vs_fp32 = (float)fp32_bytes / total_bytes;
-    float compression_vs_fp16 = (float)fp16_bytes / total_bytes;
-    
-    std::cout << "  Storage: " << total_bytes << " bytes/vector "
-              << "(" << compression_vs_fp32 << "x vs FP32, "
-              << compression_vs_fp16 << "x vs FP16)\n";
-    
-    require(total_bytes == 76, "total storage should be 76 bytes per vector");
-    require(compression_vs_fp32 > 6.0f, "compression ratio vs FP32 should be > 6x");
-}
-
-void test_qjl_encode_decode_roundtrip() {
-    std::mt19937 rng(777);
-    std::normal_distribution<float> dist(0.0f, 0.1f);
-    
-    std::vector<float> matrix(kDim * QJL_PROJ_DIM);
-    qjl_generate_projection_matrix(matrix.data(), kDim, 0xDEADBEEF);
-    
-    for (int t = 0; t < 50; t++) {
-        std::vector<float> residual(kDim);
-        for (float& v : residual) v = dist(rng);
-        
-        std::vector<uint8_t> signs(QJL_BYTES_PER_VECTOR, 0);
-        float scale = qjl_encode_residual(residual.data(), matrix.data(), signs.data(), kDim);
-        
-        std::vector<float> decoded(kDim, 0.0f);
-        qjl_decode_residual(signs.data(), matrix.data(), scale, decoded.data(), kDim);
-        
-        float cos = cosine_similarity(residual, decoded);
-        // 1-bit QJL preserves direction reasonably well
-        require(cos > 0.3f || scale < 1e-6f,
-            "QJL decode should preserve direction (cosine > 0.3)");
-    }
-}
-
-}  // namespace
-
-// ── Main ───────────────────────────────────────────────────────────────
-
-int main() {
-    struct TestCase {
-        const char* name;
-        void (*fn)();
-    };
-    
-    TestCase tests[] = {
-        {"QJL zero vector",               test_qjl_zero_vector},
-        {"QJL improves over PolarQuant",  test_qjl_improves_over_polar_alone},
-        {"QJL cosine similarity gate",    test_qjl_cosine_similarity_gate},
-        {"QJL error bounds gate",         test_qjl_error_bounds_gate},
-        {"QJL deterministic",             test_qjl_deterministic},
-        {"QJL projection matrix props",   test_qjl_projection_matrix_properties},
-        {"QJL compression ratio",         test_qjl_compression_ratio},
-        {"QJL encode/decode roundtrip",   test_qjl_encode_decode_roundtrip},
-    };
-    
-    int passed = 0, failed = 0;
-    
-    std::cout << "QJL Accuracy Gate Tests (Issue #66)\n";
-    std::cout << "====================================\n\n";
-    
-    for (auto& tc : tests) {
-        std::cout << "[" << (passed + failed + 1) << "] " << tc.name << " ... ";
-        try {
-            tc.fn();
-            std::cout << "PASS\n";
-            passed++;
-        } catch (const std::exception& e) {
-            std::cout << "FAIL: " << e.what() << "\n";
-            failed++;
-        }
-    }
-    
-    std::cout << "\n====================================\n";
-    std::cout << "Results: " << passed << " passed, " << failed << " failed\n";
-    
-    return failed > 0 ? 1 : 0;
-}

tests/test_dflash_apple_silicon.py

@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+"""Tests for Apple Silicon DFlash benchmark planning helpers (issue #152)."""
+
+import os
+import sys
+from unittest.mock import patch
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(__file__)))
+
+from benchmarks.dflash_apple_silicon import (  # noqa: E402
+    build_mlx_benchmark_command,
+    detect_total_memory_gb,
+    render_report_template,
+    select_pair,
+)
+
+
+class TestPairSelection:
+    def test_prefers_qwen35_9b_on_36gb_mac(self):
+        pair = select_pair(total_memory_gb=36)
+        assert pair.slug == "qwen35-9b"
+        assert pair.base_model == "Qwen/Qwen3.5-9B"
+        assert pair.draft_model == "z-lab/Qwen3.5-9B-DFlash"
+
+    def test_falls_back_to_4b_when_memory_is_tight(self):
+        pair = select_pair(total_memory_gb=20)
+        assert pair.slug == "qwen35-4b"
+        assert pair.base_model == "Qwen/Qwen3.5-4B"
+
+
+class TestCommandGeneration:
+    def test_builds_upstream_mlx_benchmark_command(self):
+        pair = select_pair(total_memory_gb=36)
+        command = build_mlx_benchmark_command(pair, dataset="gsm8k", max_samples=64)
+        assert "python -m dflash.benchmark --backend mlx" in command
+        assert "--model Qwen/Qwen3.5-9B" in command
+        assert "--draft-model z-lab/Qwen3.5-9B-DFlash" in command
+        assert "--dataset gsm8k" in command
+        assert "--max-samples 64" in command
+        assert "--draft-sliding-window-size 4096" in command
+
+
+class TestReportTemplate:
+    def test_report_template_mentions_baseline_and_verdict(self):
+        pair = select_pair(total_memory_gb=36)
+        report = render_report_template(machine_label="M3 Max 36GB", pair=pair)
+        assert "DFlash Apple Silicon Benchmark Report" in report
+        assert "M3 Max 36GB" in report
+        assert "Qwen/Qwen3.5-9B" in report
+        assert "plain MLX or llama.cpp speculative decoding" in report
+        assert "Worth operationalizing locally?" in report
+
+
+class TestMemoryDetection:
+    @patch("benchmarks.dflash_apple_silicon.platform.system", return_value="Darwin")
+    @patch("benchmarks.dflash_apple_silicon.subprocess.check_output", return_value=b"38654705664\n")
+    def test_detect_total_memory_gb_on_macos(self, _mock_sysctl, _mock_system):
+        assert detect_total_memory_gb() == 36.0

tests/test_quant_selector.py

@@ -0,0 +1,177 @@
+#!/usr/bin/env python3
+"""Tests for quant_selector.py"""
+
+import sys
+import os
+import pytest
+from unittest.mock import patch, MagicMock
+
+sys.path.insert(0, os.path.dirname(os.path.dirname(__file__)))
+from evolution.quant_selector import (
+    QuantLevel,
+    HardwareInfo,
+    QUANT_LEVELS,
+    detect_hardware,
+    estimate_kv_cache_gb,
+    estimate_model_memory_gb,
+    select_quant_level,
+)
+
+
+class TestQuantLevels:
+    def test_levels_keep_turboquant_quality_order_with_q4_fallback_last(self):
+        """TurboQuant levels should lead, with q4_0 reserved as the non-Turbo fallback."""
+        names = [level.name for level in QUANT_LEVELS]
+        assert names[:3] == ["turbo4", "turbo3", "turbo2"]
+        assert names[-1] == "q4_0"
+
+    def test_all_levels_have_required_fields(self):
+        for level in QUANT_LEVELS:
+            assert level.name
+            assert level.bits_per_channel > 0
+            assert level.compression_ratio > 1
+            assert level.quality_label
+            assert level.layer_adaptive >= 0
+            assert level.kv_type
+
+
+class TestKVEstimate:
+    def test_basic_estimate(self):
+        # 48 layers, 8 heads, 128 dim, 32K context, 3.5 bits
+        kv_gb = estimate_kv_cache_gb(32768, 48, 8, 128, 3.5)
+        assert kv_gb > 0
+        assert kv_gb < 10  # Should be reasonable
+
+    def test_longer_context_larger(self):
+        kv_32k = estimate_kv_cache_gb(32768, 48, 8, 128, 3.5)
+        kv_128k = estimate_kv_cache_gb(131072, 48, 8, 128, 3.5)
+        assert kv_128k > kv_32k
+
+    def test_higher_bits_larger(self):
+        kv_4b = estimate_kv_cache_gb(32768, 48, 8, 128, 4.0)
+        kv_2b = estimate_kv_cache_gb(32768, 48, 8, 128, 2.0)
+        assert kv_4b > kv_2b
+
+
+class TestHardwareDetection:
+    def test_detect_returns_info(self):
+        hw = detect_hardware()
+        assert hw.total_memory_gb > 0
+        assert hw.available_memory_gb > 0
+        assert hw.detection_method
+
+    @patch("evolution.quant_selector.platform.system", return_value="Linux")
+    @patch("builtins.open", create=True)
+    def test_linux_detection(self, mock_open, mock_system):
+        mock_open.return_value.__enter__().read.return_value = (
+            "MemTotal:       32000000 kB\n"
+            "MemAvailable:   24000000 kB\n"
+        )
+        hw = _detect_linux_fallback()
+        assert hw.total_memory_gb > 20
+
+
+def _detect_linux_fallback():
+    """Helper to test Linux detection with mocked /proc/meminfo."""
+    from evolution.quant_selector import _detect_linux
+    return _detect_linux()
+
+
+class TestSelection:
+    def test_selects_turbo4_for_large_memory(self):
+        """With plenty of memory, should pick turbo4 (best quality)."""
+        with patch("evolution.quant_selector.detect_hardware") as mock_hw:
+            mock_hw.return_value = HardwareInfo(
+                total_memory_gb=64,
+                available_memory_gb=48,
+                gpu_memory_gb=64,
+                gpu_name="Test GPU",
+                cpu_cores=16,
+                detection_method="mock",
+            )
+            sel = select_quant_level(model_size_gb=14.0, context_length=32768)
+            assert sel.level.name == "turbo4"
+            assert sel.headroom_gb > 0
+
+    def test_selects_smaller_for_tight_memory(self):
+        """With tight memory, should pick a smaller quant."""
+        with patch("evolution.quant_selector.detect_hardware") as mock_hw:
+            mock_hw.return_value = HardwareInfo(
+                total_memory_gb=16,
+                available_memory_gb=12,
+                gpu_memory_gb=16,
+                gpu_name="Test GPU",
+                cpu_cores=8,
+                detection_method="mock",
+            )
+            sel = select_quant_level(model_size_gb=14.0, context_length=131072)
+            # Should pick a smaller quant for 128K context on 16GB
+            assert sel.level.bits_per_channel <= 4.0
+
+    def test_preferred_level(self):
+        """User can force a specific level."""
+        with patch("evolution.quant_selector.detect_hardware") as mock_hw:
+            mock_hw.return_value = HardwareInfo(
+                total_memory_gb=64,
+                available_memory_gb=48,
+                cpu_cores=16,
+                detection_method="mock",
+            )
+            sel = select_quant_level(
+                model_size_gb=14.0, context_length=32768,
+                preferred_level="turbo2"
+            )
+            assert sel.level.name == "turbo2"
+
+    def test_env_vars_populated(self):
+        with patch("evolution.quant_selector.detect_hardware") as mock_hw:
+            mock_hw.return_value = HardwareInfo(
+                total_memory_gb=64,
+                available_memory_gb=48,
+                cpu_cores=16,
+                detection_method="mock",
+            )
+            sel = select_quant_level(model_size_gb=14.0, context_length=32768)
+            assert "TURBO_LAYER_ADAPTIVE" in sel.env_vars
+            assert "-ctk" in sel.server_flags
+            assert "-ctv" in sel.server_flags
+
+    def test_warnings_on_low_headroom(self):
+        with patch("evolution.quant_selector.detect_hardware") as mock_hw:
+            mock_hw.return_value = HardwareInfo(
+                total_memory_gb=18,
+                available_memory_gb=14,
+                gpu_memory_gb=18,
+                gpu_name="Test GPU",
+                cpu_cores=8,
+                detection_method="mock",
+            )
+            sel = select_quant_level(model_size_gb=16.0, context_length=65536)
+            assert len(sel.warnings) > 0
+
+    def test_falls_back_to_turbo2_when_nothing_fits(self):
+        with patch("evolution.quant_selector.detect_hardware") as mock_hw:
+            mock_hw.return_value = HardwareInfo(
+                total_memory_gb=8,
+                available_memory_gb=6,
+                gpu_memory_gb=8,
+                gpu_name="Tiny GPU",
+                cpu_cores=4,
+                detection_method="mock",
+            )
+            sel = select_quant_level(model_size_gb=16.0, context_length=131072)
+            assert sel.level.name == "turbo2"
+
+    def test_reasoning_contains_key_info(self):
+        with patch("evolution.quant_selector.detect_hardware") as mock_hw:
+            mock_hw.return_value = HardwareInfo(
+                total_memory_gb=32,
+                available_memory_gb=24,
+                is_apple_silicon=True,
+                chip_name="M4 Max",
+                cpu_cores=16,
+                detection_method="mock",
+            )
+            sel = select_quant_level(model_size_gb=14.0, context_length=32768)
+            assert "turbo4" in sel.reasoning
+            assert "M4 Max" in sel.reasoning or "32GB" in sel.reasoning

tests/test_tool_call_integration.py

@@ -0,0 +1,338 @@
+"""
+Integration test: turboquant compressed model passes hermes tool calls (issue #82).
+
+Validates that a TurboQuant-compressed model can:
+1. Parse hermes tool schemas correctly
+2. Format tool calls in OpenAI-compatible format
+3. Pass through the hermes agent conversation loop
+
+Tests are structured as contract tests -- they validate the schema/format
+compatibility without requiring a running model server. The live inference
+test is skipped by default (requires llama-server with TurboQuant model).
+
+Usage:
+    pytest tests/test_tool_call_integration.py -v
+    pytest tests/test_tool_call_integration.py -v -k live  # run live test if server available
+"""
+import json
+import os
+import pathlib
+import re
+import unittest
+
+import pytest
+
+ROOT = pathlib.Path(__file__).resolve().parents[1]
+PROFILE_PATH = ROOT / "profiles" / "hermes-profile-gemma4-turboquant.yaml"
+BENCHMARKS_DIR = ROOT / "benchmarks"
+
+
+class TestHermesProfileSchema(unittest.TestCase):
+    """Validate the hermes profile YAML has required fields for tool calling."""
+
+    @classmethod
+    def setUpClass(cls):
+        import yaml
+        cls.profile = yaml.safe_load(PROFILE_PATH.read_text())
+
+    def test_profile_has_providers(self):
+        assert "providers" in self.profile, "Profile must define providers"
+        assert "primary" in self.profile["providers"], "Must have primary provider"
+
+    def test_primary_provider_has_endpoint(self):
+        primary = self.profile["providers"]["primary"]
+        assert "endpoint" in primary, "Primary provider must have endpoint"
+        assert primary["endpoint"].startswith("http"), "Endpoint must be HTTP(S) URL"
+
+    def test_primary_provider_has_api_path(self):
+        primary = self.profile["providers"]["primary"]
+        assert "api_path" in primary, "Primary provider must have api_path"
+        assert "/chat/completions" in primary["api_path"], (
+            "api_path should be OpenAI-compatible /chat/completions"
+        )
+
+    def test_turboquant_settings_present(self):
+        primary = self.profile["providers"]["primary"]
+        assert "turboquant" in primary, "Must have turboquant config section"
+        tq = primary["turboquant"]
+        assert tq.get("enabled") is True, "TurboQuant must be enabled"
+        assert tq.get("kv_type") in ("turbo2", "turbo3", "turbo4"), (
+            "kv_type must be turbo2, turbo3, or turbo4"
+        )
+
+    def test_context_window_configured(self):
+        primary = self.profile["providers"]["primary"]
+        assert "context" in primary, "Must have context config"
+        ctx = primary["context"]
+        assert ctx.get("max_tokens", 0) >= 8192, (
+            "max_tokens should be >= 8192 for TurboQuant value proposition"
+        )
+
+
+class TestToolSchemaCompatibility(unittest.TestCase):
+    """Verify hermes tool schemas serialize to valid JSON for OpenAI tool_calls."""
+
+    SAMPLE_TOOL_SCHEMAS = [
+        {
+            "type": "function",
+            "function": {
+                "name": "read_file",
+                "description": "Read a text file with line numbers.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "path": {"type": "string", "description": "File path"},
+                        "offset": {"type": "integer", "default": 1},
+                        "limit": {"type": "integer", "default": 500},
+                    },
+                    "required": ["path"],
+                },
+            },
+        },
+        {
+            "type": "function",
+            "function": {
+                "name": "execute_code",
+                "description": "Run a Python script.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "code": {"type": "string", "description": "Python code"},
+                    },
+                    "required": ["code"],
+                },
+            },
+        },
+        {
+            "type": "function",
+            "function": {
+                "name": "web_search",
+                "description": "Search the web.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "query": {"type": "string"},
+                        "max_results": {"type": "integer", "default": 5},
+                    },
+                    "required": ["query"],
+                },
+            },
+        },
+    ]
+
+    def test_tool_schemas_serialize_to_json(self):
+        """Tool schemas must serialize without errors."""
+        serialized = json.dumps(self.SAMPLE_TOOL_SCHEMAS)
+        assert len(serialized) > 0
+        parsed = json.loads(serialized)
+        assert len(parsed) == len(self.SAMPLE_TOOL_SCHEMAS)
+
+    def test_tool_schemas_have_required_openai_fields(self):
+        """Each tool schema must have the fields OpenAI expects."""
+        for tool in self.SAMPLE_TOOL_SCHEMAS:
+            assert tool["type"] == "function", "Tool type must be 'function'"
+            fn = tool["function"]
+            assert "name" in fn, "Function must have name"
+            assert "description" in fn, "Function must have description"
+            assert "parameters" in fn, "Function must have parameters"
+            params = fn["parameters"]
+            assert params["type"] == "object", "Parameters type must be 'object'"
+            assert "properties" in params, "Parameters must have properties"
+
+    def test_tool_call_response_format(self):
+        """Verify tool_call response matches OpenAI format."""
+        tool_call = {
+            "id": "call_abc123",
+            "type": "function",
+            "function": {
+                "name": "read_file",
+                "arguments": json.dumps({"path": "/tmp/test.txt"}),
+            },
+        }
+        args = json.loads(tool_call["function"]["arguments"])
+        assert args["path"] == "/tmp/test.txt"
+        assert tool_call["function"]["name"] in [
+            t["function"]["name"] for t in self.SAMPLE_TOOL_SCHEMAS
+        ]
+
+    def test_tool_names_are_valid_identifiers(self):
+        """Tool names must be valid Python identifiers for hermes dispatch."""
+        for tool in self.SAMPLE_TOOL_SCHEMAS:
+            name = tool["function"]["name"]
+            assert re.match(r"^[a-zA-Z_][a-zA-Z0-9_]*$", name), (
+                f"Tool name \'{name}\' is not a valid identifier"
+            )
+
+
+class TestTurboquantServerConfig(unittest.TestCase):
+    """Validate server startup configuration matches hermes profile."""
+
+    def test_server_command_has_turboquant_flags(self):
+        """The server command in the profile must include -ctk/-ctv flags."""
+        profile_text = PROFILE_PATH.read_text()
+        assert "-ctk" in profile_text, "Profile server command must include -ctk flag"
+        assert "-ctv" in profile_text, "Profile server command must include -ctv flag"
+
+    def test_server_command_has_context_flag(self):
+        """Server command must set context size."""
+        profile_text = PROFILE_PATH.read_text()
+        assert re.search(r"-c\s+\d+", profile_text), (
+            "Server command must include -c <context_size> flag"
+        )
+
+    def test_layer_adaptive_env_var(self):
+        """Profile must set TURBO_LAYER_ADAPTIVE env var."""
+        profile_text = PROFILE_PATH.read_text()
+        assert "TURBO_LAYER_ADAPTIVE" in profile_text, (
+            "Profile must configure TURBO_LAYER_ADAPTIVE"
+        )
+
+
+class TestBenchmarkData(unittest.TestCase):
+    """Validate benchmark test prompts include tool-call test cases."""
+
+    @classmethod
+    def setUpClass(cls):
+        prompts_path = BENCHMARKS_DIR / "test_prompts.json"
+        cls.prompts = json.loads(prompts_path.read_text())
+
+    def test_has_tool_call_test_prompt(self):
+        """Benchmark prompts must include a tool-call format test."""
+        categories = [p.get("category") for p in self.prompts]
+        assert "tool_call_format" in categories, (
+            "Benchmark must include a tool_call_format test case"
+        )
+
+    def test_tool_call_prompt_expects_json(self):
+        """Tool call test prompt must expect JSON in the response."""
+        tool_prompt = next(
+            p for p in self.prompts if p.get("category") == "tool_call_format"
+        )
+        pattern = tool_prompt.get("expected_pattern", "")
+        assert "json" in pattern.lower() or "\\{" in pattern, (
+            "Tool call prompt must expect JSON-formatted response"
+        )
+
+
+@pytest.mark.skipif(
+    not os.environ.get("TURBOQUANT_SERVER_URL"),
+    reason="No TurboQuant server available (set TURBOQUANT_SERVER_URL to run)",
+)
+class TestLiveToolCallIntegration:
+    """Live integration test -- requires running llama-server with TurboQuant."""
+
+    def test_server_health(self):
+        """Server must respond to /v1/models endpoint."""
+        import requests
+        url = os.environ["TURBOQUANT_SERVER_URL"]
+        resp = requests.get(f"{url}/v1/models", timeout=10)
+        assert resp.status_code == 200
+        data = resp.json()
+        assert "data" in data
+        assert len(data["data"]) > 0
+
+    def test_tool_call_completion(self):
+        """Model must return a valid tool_call for a read_file prompt."""
+        import requests
+        url = os.environ["TURBOQUANT_SERVER_URL"]
+        tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "read_file",
+                    "description": "Read a file",
+                    "parameters": {
+                        "type": "object",
+                        "properties": {"path": {"type": "string"}},
+                        "required": ["path"],
+                    },
+                },
+            }
+        ]
+        resp = requests.post(
+            f"{url}/v1/chat/completions",
+            json={
+                "model": "gemma-4",
+                "messages": [
+                    {"role": "user", "content": "Read the file at /tmp/test.txt"}
+                ],
+                "tools": tools,
+                "tool_choice": "auto",
+            },
+            timeout=120,
+        )
+        assert resp.status_code == 200
+        data = resp.json()
+        choice = data["choices"][0]
+        msg = choice["message"]
+        if "tool_calls" in msg and msg["tool_calls"]:
+            tc = msg["tool_calls"][0]
+            assert tc["type"] == "function"
+            assert tc["function"]["name"] == "read_file"
+            args = json.loads(tc["function"]["arguments"])
+            assert "path" in args
+        else:
+            assert len(msg.get("content", "")) > 0
+
+    def test_tool_call_with_multiple_tools(self):
+        """Model must handle multiple available tools."""
+        import requests
+        url = os.environ["TURBOQUANT_SERVER_URL"]
+        tools = [
+            {
+                "type": "function",
+                "function": {
+                    "name": "read_file",
+                    "description": "Read a file",
+                    "parameters": {
+                        "type": "object",
+                        "properties": {"path": {"type": "string"}},
+                        "required": ["path"],
+                    },
+                },
+            },
+            {
+                "type": "function",
+                "function": {
+                    "name": "web_search",
+                    "description": "Search the web",
+                    "parameters": {
+                        "type": "object",
+                        "properties": {"query": {"type": "string"}},
+                        "required": ["query"],
+                    },
+                },
+            },
+            {
+                "type": "function",
+                "function": {
+                    "name": "execute_code",
+                    "description": "Run Python code",
+                    "parameters": {
+                        "type": "object",
+                        "properties": {"code": {"type": "string"}},
+                        "required": ["code"],
+                    },
+                },
+            },
+        ]
+        resp = requests.post(
+            f"{url}/v1/chat/completions",
+            json={
+                "model": "gemma-4",
+                "messages": [
+                    {"role": "user", "content": "Search the web for 'bitcoin price'"}
+                ],
+                "tools": tools,
+                "tool_choice": "auto",
+            },
+            timeout=120,
+        )
+        assert resp.status_code == 200
+        data = resp.json()
+        assert "choices" in data
+        assert len(data["choices"]) > 0
+
+
+if __name__ == "__main__":
+    unittest.main()