phase 8: add benchmark framework + baseline results

- bench-gpt2 binary: runs 50 prompts, measures TTFT/TBT per prompt, outputs JSON
- bench_compare.py: compares xserv vs transformers token-by-token + timing
- Baseline results: 50/50 correctness, 400ms TTFT / 407ms TBT (100x slower than PyTorch)
- Bottlenecks documented: no KV cache, CPU round-trips, cuBLAS handle churn

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

docs/benchmarks/phase8-gpt2-baseline.md

@@ -0,0 +1,35 @@
+# Phase 8 Benchmark: GPT-2 124M Baseline
+
+**Date**: 2026-05-21
+**Hardware**: RTX 5090 (32GB, CC 12.0, 170 SMs)
+**Model**: GPT-2 124M (FP32)
+**Config**: 50 prompts × 20 generated tokens, greedy decoding, no KV cache
+
+## Correctness
+
+| Metric | Result |
+|--------|--------|
+| Prompts tested | 50 |
+| Token-level match vs transformers | **50/50 (100.0%)** |
+| Mismatches | 0 |
+
+## Performance
+
+| Metric | xserv | transformers (PyTorch) | Ratio |
+|--------|-------|----------------------|-------|
+| TTFT (avg) | 400.6 ms | 4.0 ms | 100x slower |
+| TBT (avg) | 407.2 ms | 3.8 ms | 106x slower |
+| Throughput | 2.5 tok/s | 260 tok/s | 0.01x |
+
+## Known Bottlenecks
+
+1. **No KV Cache**: full recompute per token (O(S²) attention every step)
+2. **CPU round-trips**: ~100 GPU→CPU→GPU transfers per forward pass for add/bias/split_qkv/merge_heads
+3. **cuBLAS handle per matmul**: ~50 handle create/destroy per forward pass
+4. **No kernel fusion**: every op is a separate kernel launch + sync
+
+## Tracking
+
+| Phase | TTFT (ms) | TBT (ms) | tok/s | Correctness | Notes |
+|-------|-----------|----------|-------|-------------|-------|
+| 8 (baseline) | 400.6 | 407.2 | 2.5 | 50/50 | No KV cache, CPU round-trips |