Feat/silicon path runtime - #1

Merged

sriharshapy merged 86 commits into

feat/silicon-path-runtime

Aug 15, 2026

Commits on Aug 10, 2026

runtime: the batch wire format, with no field for an address
sriharshapy
committed
2fd5ba1
runtime: the IDL, qaic generation, and the MIT attribution it obliges
sriharshapy
committed
ddded37
runtime: a test that fails if qaic's fail-closed check is removed
sriharshapy
committed
8bd50ef
runtime: generate DSP entry points from the spec that already declares them
sriharshapy
committed
3d9c0b6
runtime: make the requires tests distinguish checked from documented
sriharshapy
committed
c2ce470
skel: the host writes fds and offsets, the DSP writes addresses
sriharshapy
committed
8980275
skel: test the clear-before-lookup invariant where it lives, not by file order
sriharshapy
committed
6430ca9
skel: VTCM size from the runtime, and the reclaim callback registered
sriharshapy
committed
30a0cb3
skel: test that every VTCM failure path returns a status
sriharshapy
committed
05dcaca
skel: batch walk, PCYCLE around the kernel call only, failure as the default
sriharshapy
committed
3d8c14e
runtime: the simulator host and the qexe link, from a recipe known to work
sriharshapy
committed
a281603
runtime: say plainly what the simulator path does not prove
sriharshapy
committed
09d19f6
host: dlopen the driver, unsigned PD on CDSP, and no address on the wire
sriharshapy
committed
938a8f2
host: tighten the URI, dlopen and request-id tests to check use, not co-occurrence
sriharshapy
committed
d38e04f
runtime: a QuRT-hosted sim build, so VTCM acquisition is really exercised

sriharshapy
and
claude
committed
0dfb41c
docs: retarget Task 8 and the spec at the QuRT-hosted sim build

sriharshapy
and
claude
committed
c67654c
docs: mark the standalone qexe link recipe superseded, with the reason

sriharshapy
and
claude
committed
c5016e6
runtime: actually git-ignore the generated sim configs, and say where writes land

sriharshapy
and
claude
committed
ccae3fb
chore: untrack two simulator artifacts I committed by accident

sriharshapy
and
claude
committed
d0ef88c

Commits on Aug 11, 2026

stage 1 GREEN: scale_fp16 through the skel batch path on the simulator, and an unmapped fd refused
sriharshapy
committed
0a6225e
stage 2: cross-compile hexlib_run for arm64 and the skel as a device .so
sriharshapy
committed
2920c5f
qdc: submit, and detect completion from log files rather than job status
sriharshapy
committed
5fa5203
qdc: use the SDK's own API-key client, and stop requiring a base URL we can default
sriharshapy
committed
df3a8bf
host: define the remote_handle64 forwarders, so the driver stays dlopen'd
sriharshapy
committed
897f4ac
tests: one comment-aware block matcher, so 40 source assertions inspect what they claim
sriharshapy
committed
e3bc6f7
qdc: the on-device test, and --device on the CLI
sriharshapy
committed
765d786
host: --unmapped, --coherency-check, and print the DSP's cycle count
sriharshapy
committed
f9edc3b
docs: the handoff record said three flags did not exist -- they landed hours ago

sriharshapy
and
claude
committed
ba5762a
fix: the arch check could never pass, coherency misfired on -0.0, and a failed job exited 0

sriharshapy
and
claude
committed
8b35c00
tests: prove the -0.0 coherency fix behaviourally, and stop skipping silently
sriharshapy
committed
63becd3
tests: the slicer was comment-aware only at its boundaries, so payloads were blind

sriharshapy
and
claude
committed
9af5803
tests: CI ran zero tests and exited 2, and the guard against that was blind to it

sriharshapy
and
claude
committed
07d0cee
docs: the HVX-clean rule banned what ATTRIBUTION.md deliberately does

sriharshapy
and
claude
committed
f864ead
runtime: requires was enforced on neither side of the DSP transport

sriharshapy
and
claude
committed
7af6cd3
docs: rewrite the front door, and fix three numbers that were wrong in public

sriharshapy
and
claude
committed
853ef57
ci: the `|| pip install pytest` fallback could never fire, so pytest was never installed

sriharshapy
and
claude
committed
eb8ab1e
runtime: the DSP's cycle read had no handling of the one bit it depends on

sriharshapy
and
claude
committed
4d64ee2
skel: one contended VTCM page would have failed every session on silicon

sriharshapy
and
claude
committed
ddb068d
skel: hwinfo claimed five DSP facts and reports one

sriharshapy
and
claude
committed
e35fb30
tests: an "outputs first" refactor would have swapped every kernel's pointers silently

sriharshapy
and
claude
committed
e346808
qdc: a 3-minute budget would submit a 240-minute job, and a skipped test counted as a pass

sriharshapy
and
claude
committed
d3f39ce
qdc: the traversal guard only held on the OS you happened to test on

sriharshapy
and
claude
committed
3e2d362
docs: the address claim was false for base, and three published figures overstated

sriharshapy
and
claude
committed
8b20443
tests: blanking comments closed one vehicle for C2 and left the string literal open

sriharshapy
and
claude
committed
1fe8871
exec: layernorm gated green for a day without being dispatchable at all

sriharshapy
and
claude
committed
67766e4
fix: the four remaining Importants from the merge gate

sriharshapy
and
claude
committed
a7f586f
test: the simhost fd-patch bound, checked on the simulator rather than by grep

sriharshapy
and
claude
committed
ca41626
test: the simhost bound test was green with the bound removed

sriharshapy
and
claude
committed
29aa165
runtime: the block-quantized weight path, so matmul_epilogue can exist

sriharshapy
and
claude
committed
dd0ca96
kernels: transpose_hd_fp16, the perm(0,2,1) the sibling kernel called hard

sriharshapy
and
claude
committed
fd0c1ca
kernels: softmax_fp16 and rope_2d_fp16, both vectorised and both discriminating

sriharshapy
and
claude
committed
ed81390
runtime: one op kind, several kernels -- SPECS keyed by variant

sriharshapy
and
claude
committed
27eae11
exec: softmax and rope_2d dispatch -- coverage 111 -> 159 of 259

sriharshapy
and
claude
committed
922f29f
kernels: patchify_fp32, and `merge` is not metadata

sriharshapy
and
claude
committed
b58c207
test: the four new kernels actually dispatch, and _out_shape had to learn patchify

sriharshapy
and
claude
committed
0e14272
exec: Scalar.codes, so a string attr can cross a wire that carries only numbers

sriharshapy
and
claude
committed
9568dae
fix: eight of the merge gate's Minors, and one of them was wrong

sriharshapy
and
claude
committed
5085054
exec: a reference q4_0 quantizer, because nothing in the tree could make one

sriharshapy
and
claude
committed
f6d1bd0
test: the whole encoder end to end, every kernel on the simulator's DSP path

sriharshapy
and
claude
committed
62edbab

Commits on Aug 12, 2026

kernels: matmul_epilogue_fp16, q4_0 weights through bias and gelu

sriharshapy
and
claude
committed
bff4d66
fix: the gate's simulator ceiling was below what a scalar near-miss costs

sriharshapy
and
claude
committed
9a2c99d
gate: attribute a simulator timeout instead of guessing at it

sriharshapy
and
claude
committed
23e3049
kernels: matmul_fp16, the encoder's 24 attention matmuls

sriharshapy
and
claude
committed
7a41e96
qdc: the first two reasons a device job could never report anything

sriharshapy
and
claude
committed
18b4e14
qdc: the tests run on the runner and reach the phone by adb, not on the phone

sriharshapy
and
claude
committed
a491774
qdc: a flat zip is accepted, dispatched, and never runs

sriharshapy
and
claude
committed
94e2def
exec: matmul dispatches -- 160 of 259 encoder ops becomes 184
sriharshapy
committed
964ee1a
exec: matmul_epilogue dispatches -- q4_0 weights across the wire
sriharshapy
committed
7094ac8
test: the encoder runs with no reference fallback at all

sriharshapy
and
claude
committed
03ccd9c
test: the encoder's coverage number, asked of select() not of a tally
sriharshapy
committed
da83304
fix: matmul_fp16 read B rows aligned, and only a multiple-of-64 N made that true
sriharshapy
committed
cf25017
fix: two claims made today that measurement did not support

sriharshapy
and
claude
committed
9ecaf15
fix: gate matmul_fp16 at N=200, the shape the aligned-load bug hid behind

sriharshapy
and
claude
committed
06cc2a4
qdc: the first path that reaches silicon, and what it found there

sriharshapy
and
claude
committed
91ad9de

Commits on Aug 13, 2026

kernels: patchify_fp32's RESULT.md timestamp, from a re-measurement

sriharshapy
and
claude
committed
62f67eb
docs: untrack docs/ entirely, and fix the one test that depended on it

sriharshapy
and
claude
committed
524a41f
exec: the whole encoder as ONE batch -- a single entry into the DSP, one exit

sriharshapy
and
claude
committed
30479e1
fix: the skel was linked under a name FastRPC never looks for

sriharshapy
and
claude
committed
5677167
host: --batch can stage an in-out arena, which is what a whole plan is

sriharshapy
and
claude
committed
1e7fc54
skel: write the DSP's cache back, without which the host reads stale data

sriharshapy
and
claude
committed
ab383c5
qdc: the session can run a whole plan, and the dual-name push is gone

sriharshapy
and
claude
committed
3089aa9
models: the checkpoint->const mapping moves into the library

sriharshapy
and
claude
committed
ab9d463
quant: q8_0, because four bits is not enough for this encoder

sriharshapy
and
claude
committed
da76dcc
hmx: the first HMX kernel, and the wall it hit -- DOES NOT PASS ITS GATE

sriharshapy
and
claude
committed
da24dd3
hmx: the four things HMX needs before mxmem runs, read out of llama.cpp

sriharshapy
and
claude
committed
c830c5a

Commits on Aug 15, 2026

docs: the tracked documentation catches up with ten commits of measurement

sriharshapy
and
claude
committed
7586b36