[thanhdq1o2]

What

• docs/benchmark/qualcomm-iq10.md, "Q8_0 weights":
  • every Q8_0 configuration runs six checks against the CPU-only build at F32: fixed inputs and five served LIBERO observations (obs0 to obs4);
  • --keep-float lists searched to pass all six checks;
  • the configuration per model, with its quantize_gguf.py flags;
  • per model, a table of every configuration with its six checks and a table of every search with its outcome, in place of the single 62-row table;
  • levels named by what they put in Q8_0: base, base + vision, base + vision + DiT.
• docs/backend/hexagon.md: the Q8_0 summary and example follow the six-check results.
• CHANGELOG.md.

Why

Lists picked on one observation fail on others: OpenVLA-OFT's obs0 list passes obs0 (2.5e-3) and fails obs1 (4.3e-3). The section now reports each configuration on six checks and picks, per model, the fastest one that passes all of them and has a recorded --keep-float list.

Results on the IQ-10 at 71cdb09, mean latency and System RAM under the TEMPLATE protocol.

• base puts in Q8_0 the 2-D weights outside the vision tower and the action head; base + vision adds the vision tower; n kept F16 keeps n of those weights out of Q8_0 with --keep-float.
• Weights in Q8_0 counts the weights in Q8_0 out of those the level can put in Q8_0.
• For each check, max|Δ| divided by its threshold (2.9e-3, or 1.1 times F16 where F16 is above 2.9e-3) gives a ratio; a configuration's score is its largest ratio, and it passes all checks at 1 or less.
• A model fails when no configuration passes; its row gives the best score its searches reached, or, where no search ran, the score with all of base in Q8_0.

Model	Q8_0 result	Q8_0 configuration	Weights in Q8_0	F16 ms	Q8_0 ms	vs F16	F16 System RAM MiB	Q8_0 System RAM MiB	vs F16
π0	passes all checks	base + vision, 8 kept F16	281 of 289	1123.3	937.7	-16.5%	5662	3326	-41.3%
π0.5	passes all checks	base + vision, 181 kept F16	108 of 289	1107.8	1061.2	-4.2%	5980	5095	-14.8%
Octo-Small	passes all checks	base + vision, 4 kept F16	119 of 123	84.1	81.5	-3.1%	725	407	-43.9%
ACT	passes all checks	base	35 of 35	152.5	150.9	-1.0%	469	405	-13.6%
TurboVLA	passes all checks	base, 12 kept F16	121 of 133	190.2	189.7	-0.3%	1079	881	-18.4%
Evo-1	passes all checks	base, 66 kept F16	34 of 100	1659.0	1657.4	-0.1%	1582	1494	-5.6%
OpenVLA-OFT	fails: best score 1.03	-	-	1125.2	-	-	15266	-	-
VLA-Adapter	fails: best score 1.03	-	-	669.2	-	-	3569	-	-
GR00T N1.5	fails: best score 1.23	-	-	529.3	-	-	3659	-	-
GR00T N1.6	fails: best score 1.33	-	-	682.6	-	-	4929	-	-
GR00T N1.7	fails: best score 1.40	-	-	744.9	-	-	5108	-	-
SmolVLA	fails: not searched, score 44 with all of base in Q8_0	-	-	584.0	-	-	1226	-	-
VLA-JEPA	fails: not searched, score 36 with all of base in Q8_0	-	-	496.8	-	-	4018	-	-

Every configuration with its six checks, and every search, passing or not: Per model.

Verified

• Builds clean under -Wall -Wextra (first-party code): no code change
• ctest passes: no code change
• Numeric output unchanged: docs only

Archs and backends tested: SmolVLA, π0, π0.5, Evo-1, GR00T N1.5 / N1.6 / N1.7, VLA-JEPA, VLA-Adapter, OpenVLA-OFT, Octo-Small, TurboVLA and ACT on the IQ-10's Hexagon NPU, against the CPU-only build (-DGGML_HEXAGON=OFF, --weight-dtype f32 --flash-attn 0).