Datasets:

---

arm-bench

/

arm-bench-trace

Like 2

Follow

CMU l3 3

Summary

Real-activation workloads for the e2e llama.cpp GEMM operators of three models and MoE operators, rebuilt with kgen
(collect -> add workload -> capture -> calibrate), the operator sets re-selected with E2E total running share rule, and the quantized gemm definitions unified on the ggml block-bytes ABI under plain names (no ggml_ prefix), so that All gemm kernels supports e2e replay test.

E2E total running share rules

Identify Execution Candidates (role, type, shape):
- Filter for unique operator candidates defined by their role, data type, and tensor shape to prevent redundant benchmarking of identical kernels.
- Skip operators too small to time reliably to eliminate measurement noise.
Rank Candidates by Time Share (Not Count):
- Rank operators using estimated execution time rather than raw layer or call counts.Operators with higher time share are more critical to benchmark, as they contribute more to overall execution time.
- Feel free to customize your ranking metric based on the model/kernel's characteristics, for example: For memory-bound decode, estimate time share based on weight bytes accessed per token.For MoE experts, scale memory/compute weight by the active routing fraction per token (e.g., top-$k$ active experts).
Apply Pareto Tail Cut (90% Cumulative Share per Group):
- Group candidates by (model, quantization type).
- Within each group, rank candidates by share, calculate cumulative share relative to the group total, and retain operators until the running total first reaches or crosses 90%.
- Drop all remaining long-tail operators beyond the crossing point.
Provide Audit Trail & Context:
- List all dropped operators explicitly for transparency.
- State explicitly whether the time share is estimated or measured.

What is in this PR

The following gemm kernels are recalibrated and selected with the group:

The following moe kernels are recalibrated with model qwen1.5-moe-a2.7b: moe_q4_k_m_e60_k4_d2048_ff1536, moe_q4_k_m_e64_k8_d2048_ff1024, moe_q8_0_e60_k4_d2048_ff1408, moe_q8_0_e64_k8_d2048_ff1024, no new tensor are produced.

Removed definitions (with their workloads, tensors and every author's solutions)

The following gemms are removed based on e2e total running share rules:

• Qwen3.5-4B: gemm_q4_K_n2560_k4096, gemm_q4_K_n1024_k2560, gemm_q6_K_n1024_k2560
• Llama-3.1-8B: gemm_q4_K_n1024_k4096, gemm_q6_K_n1024_k4096

The following gemms are replaced by unified-ABI definitions:

• legacy triplet-ABI gemm_q4_k_m_n{1024_k2048,1408_k2048,2048_k1024,2048_k1536,2048_k2048} and gemm_q8_0_n{1408_k2048,2048_k1408} (replaced by the unified-ABI definitions above; the Qwen1.5-MoE shapes are re-created)
• the old gemm_ggml_* names (renamed to gemm_*)

Other files

• expected_sets_sve.json, subset_sve.json, subset_sme.json: gemm_q4_k_m_* -> gemm_q4_K_* (same shapes)
• TASKS.md: e2e section (22 operators), llama.cpp section (renamed gemms, current workload counts), summary
• bf16 gemm, moe and rms_norm workloads refreshed to the prefill M range (1..512)

How the workloads were built

• Activations: each model run in PyTorch bf16 on ShareGPT prompts (up to 512 tokens), forward-hooked per layer; OLMoE experts rebuilt as per-expert nn.Linear (verified bit-identical to transformers' reference loop) so each expert call is one gemm with M = tokens routed to it.
• Operator selection: shapes ranked by per-token weight-byte share (decode, memory-bound estimate); per quantization type, keep operators until the running share first reaches 90% of that type's total (as E2E total running share rule explained). Qwen3.5-4B and Llama-3.1-8B use the shares already recorded in their e2e inventories (e2e-share tags); OLMoE comes from a kgen discover plugin run over the GGUF headers of OLMoE-1B-7B Q4_K_M and Q8_0.
• baseline_sqnr_db calibrated on aarch64 (Graviton4) with the corrected bf16 tensor loader. The Qwen1.5-MoE definitions use deterministic random inputs.
• The Llama weights come from the ungated unsloth/Meta-Llama-3.1-8B mirror.

Notes for reviewers

• Needs the bf16 bit-pattern fix in bench/runtime/inputs.py (code repo, not part of this PR): half-precision tensor inputs are stored as raw 16-bit patterns.
• The Fable solutions of the kept operators were optimized on the earlier 6-workload (M <= 32) sets and have not been re-evaluated on these workloads.
• OLMoE operators are marked ? in TASKS.md (no recorded speedup run yet); the 4 restored Qwen1.5-MoE definitions have no e2e tag.
• LFS: 412 new tensor files (~242 MB).

Will be merged after e2e gemm kernel rerun ends