Add single-dispatch layer-by-layer multi-head attention

[andrej]

"Naive" alternative implementation for multi-head attention from the currently checked-in data-flow design. This is a simple layer-by-layer implementation, but it uses the single-dispatch mechanism to fuse it all into one MLIR file and save on CPU roundtrips and XRT overheads.

Includes two variants:

1. "core": Only does the core matmuls and softmax; assumes projected and repeated inputs Q, K, V. This matches the functionality of the checked-in dataflow MHA.
2. "projected": Performs the Q, K, V projections, applies a RoPE positional embedding and repeats K and V matrices for grouped-query attention. Takes an embedding vector and RoPE angles as input.

[andrej]

Can we reuse the reference from the existing mha? (Note: does not include RoPE and Q, K, V projections, but some code reuse should be possible.)

[github-actions]

📊 Test Results for Test Example Applications

1d87fe8 (2026_04_07_21_05_39)

IRONCLAD

Tested on 2026_04_07_21_05_39 at commit 1d87fe8.

Test	Checks	TTFT (mean)	TPS (mean)
llama_3.2_1b_prompt_1024_tokens_1	✅ 5/5	2.13	n/a
llama_3.2_1b_prompt_1024_tokens_40	✅ 5/5	2.18	4.31
llama_3.2_1b_prompt_13_tokens_1	✅ 5/5	2.09	n/a
llama_3.2_1b_prompt_13_tokens_40	✅ 5/5	2.09	4.31

📈 Trends (vs main branch) for Test Example Applications

1d87fe8 (2026_04_07_21_05_39)

IRONCLAD Trends

llama_3.2_1b

Commit/Date	Num Tokens (max)	Num Tokens (mean)	Num Tokens (median)	Num Tokens (min)	Num Tokens (stddev)	TPS (max)	TPS (mean)	TPS (median)	TPS (min)	TPS (stddev)	TTFT (max)	TTFT (mean)	TTFT (median)	TTFT (min)	TTFT (stddev)	Total (max)	Total (mean)	Total (median)	Total (min)	Total (stddev)
130b6ea — 2025-12-05 21:33:12	40.00 (+0.00%)	40.00 (+0.00%)	40.00 (+0.00%)	40.00 (+0.00%)	0.00 (n/a)	4.71 (-0.42%)	4.64 (-0.09%)	4.64 (+0.65%)	4.55 (-0.22%)	0.05 (-17.66%)	4.41 (-0.34%)	4.39 (-0.19%)	4.38 (-0.33%)	4.37 (-0.15%)	0.01 (-25.90%)	12.96 (-0.00%)	12.80 (+0.07%)	12.80 (-0.23%)	12.67 (+0.44%)	0.09 (-21.12%)
0a6c11c — 2025-12-03 23:35:15	40.00 (n/a)	40.00 (n/a)	40.00 (n/a)	40.00 (n/a)	0.00 (n/a)	4.73 (n/a)	4.64 (n/a)	4.61 (n/a)	4.56 (n/a)	0.06 (n/a)	4.42 (n/a)	4.40 (n/a)	4.40 (n/a)	4.37 (n/a)	0.02 (n/a)	12.96 (n/a)	12.79 (n/a)	12.83 (n/a)	12.62 (n/a)	0.12 (n/a)

llama_3.2_1b_prompt_1024_tokens_1

Commit/Date	TTFT (max)	TTFT (mean)	TTFT (median)	TTFT (min)	TTFT (stddev)
1d87fe8 — 2026-04-07 21:00:00	2.15 (+0.09%)	2.13 (+0.08%)	2.13 (-0.42%)	2.12 (+0.62%)	0.01 (-31.21%)
912e6bc — 2026-04-07 19:08:43	2.15 (n/a)	2.13 (n/a)	2.13 (n/a)	2.11 (n/a)	0.02 (n/a)

llama_3.2_1b_prompt_1024_tokens_40

Commit/Date	TPS (max)	TPS (mean)	TPS (median)	TPS (min)	TPS (stddev)	TTFT (max)	TTFT (mean)	TTFT (median)	TTFT (min)	TTFT (stddev)
1d87fe8 — 2026-04-07 21:00:00	4.33 (+2.90%)	4.31 (+3.44%)	4.31 (+3.58%)	4.29 (+3.77%)	0.01 (-46.93%)	2.29 (+0.48%)	2.18 (+0.83%)	2.15 (+0.80%)	2.13 (+0.61%)	0.07 (-4.73%)
912e6bc — 2026-04-07 19:08:43	4.21 (n/a)	4.17 (n/a)	4.16 (n/a)	4.14 (n/a)	0.03 (n/a)	2.28 (n/a)	2.16 (n/a)	2.13 (n/a)	2.12 (n/a)	0.07 (n/a)

llama_3.2_1b_prompt_13_tokens_1

Commit/Date	TTFT (max)	TTFT (mean)	TTFT (median)	TTFT (min)	TTFT (stddev)
1d87fe8 — 2026-04-07 21:00:00	2.10 (-0.10%)	2.09 (+0.11%)	2.09 (+0.19%)	2.09 (+0.00%)	0.01 (+8.87%)
912e6bc — 2026-04-07 19:08:43	2.10 (n/a)	2.09 (n/a)	2.09 (n/a)	2.09 (n/a)	0.01 (n/a)

llama_3.2_1b_prompt_13_tokens_40

Commit/Date	TPS (max)	TPS (mean)	TPS (median)	TPS (min)	TPS (stddev)	TTFT (max)	TTFT (mean)	TTFT (median)	TTFT (min)	TTFT (stddev)
1d87fe8 — 2026-04-07 21:00:00	4.36 (+4.23%)	4.31 (+3.57%)	4.30 (+3.44%)	4.29 (+3.23%)	0.03 (+128.30%)	2.09 (-0.38%)	2.09 (-0.04%)	2.09 (+0.00%)	2.08 (+0.44%)	0.01 (-34.93%)
912e6bc — 2026-04-07 19:08:43	4.18 (n/a)	4.16 (n/a)	4.16 (n/a)	4.15 (n/a)	0.01 (n/a)	2.10 (n/a)	2.09 (n/a)	2.09 (n/a)	2.07 (n/a)	0.01 (n/a)

llama_3.2_1b_prompt_2048_tokens_1

Commit/Date	Num_Tokens (max)	Num_Tokens (mean)	Num_Tokens (median)	Num_Tokens (min)	Num_Tokens (stddev)	TPS (max)	TPS (mean)	TPS (median)	TPS (min)	TPS (stddev)	TTFT (max)	TTFT (mean)	TTFT (median)	TTFT (min)	TTFT (stddev)
897d04e — 2026-03-06 22:56:07	1.00 (+0.00%)	1.00 (+0.00%)	1.00 (+0.00%)	1.00 (+0.00%)	0.00 (n/a)	0.00 (n/a)	0.00 (n/a)	0.00 (n/a)	0.00 (n/a)	0.00 (n/a)	2.68 (-1.06%)	2.68 (-1.06%)	2.68 (-1.06%)	2.68 (-1.06%)	0.00 (n/a)
84d3478 — 2026-02-17 23:16:23	1.00 (n/a)	1.00 (n/a)	1.00 (n/a)	1.00 (n/a)	0.00 (n/a)	0.00 (n/a)	0.00 (n/a)	0.00 (n/a)	0.00 (n/a)	0.00 (n/a)	2.70 (n/a)	2.70 (n/a)	2.70 (n/a)	2.70 (n/a)	0.00 (n/a)

llama_3.2_1b_prompt_2048_tokens_40

Commit/Date	Num_Tokens (max)	Num_Tokens (mean)	Num_Tokens (median)	Num_Tokens (min)	Num_Tokens (stddev)	TPS (max)	TPS (mean)	TPS (median)	TPS (min)	TPS (stddev)	TTFT (max)	TTFT (mean)	TTFT (median)	TTFT (min)	TTFT (stddev)
897d04e — 2026-03-06 22:56:07	40.00 (+0.00%)	40.00 (+0.00%)	40.00 (+0.00%)	40.00 (+0.00%)	0.00 (n/a)	4.00 (-1.72%)	4.00 (-1.72%)	4.00 (-1.72%)	4.00 (-1.72%)	0.00 (n/a)	2.70 (-0.44%)	2.70 (-0.44%)	2.70 (-0.44%)	2.70 (-0.44%)	0.00 (n/a)
84d3478 — 2026-02-17 23:16:23	40.00 (n/a)	40.00 (n/a)	40.00 (n/a)	40.00 (n/a)	0.00 (n/a)	4.07 (n/a)	4.07 (n/a)	4.07 (n/a)	4.07 (n/a)	0.00 (n/a)	2.71 (n/a)	2.71 (n/a)	2.71 (n/a)	2.71 (n/a)	0.00 (n/a)