Skip to content

Actions: rhCat/ds4f-disk

Actions

All workflows

Actions

Loading...
Loading

Showing runs from all workflows
75 workflow runs
75 workflow runs

Filter by Workflow

⌥ + click/return to exclude

Filter by Event

⌥ + click/return to exclude

Filter by Status

⌥ + click/return to exclude

Filter by Branch

⌥ + click/return to exclude

Filter by Actor

⌥ + click/return to exclude
attn+moe: DS4F_NO_B_MIX -- the identity-B residual update ci #75: Commit 9768441 pushed by rhCat
12s main
main
12s
attn: DS4F_V_OFFSET -- sweep the value's location in the latent ci #74: Commit 17fb880 pushed by rhCat
10s main
main
10s
attn: DEBUG9 full-v probe -- mean per-component v-std + v-mean ci #73: Commit 8ebe2c4 pushed by rhCat
17s main
main
17s
attn: DS4F_DEBUG9 -- the frozen-F probe (score spread, v-std, wmax) ci #72: Commit 9b6598e pushed by rhCat
19s main
main
19s
attn: DS4F_DEBUG8 (C/A values + Frms per layer) + DS4F_C_SCALE ci #71: Commit e3ba4b3 pushed by rhCat
10s main
main
10s
attn: DS4F_NO_SINK -- test the sink's dominance of the frozen state ci #70: Commit af88087 pushed by rhCat
12s main
main
12s
state: DS4F_STATE_RMS_TARGET -- break the dead-state fixed point ci #69: Commit d23329f pushed by rhCat
15s main
main
15s
main: DS4F_GREEDY (argmax) + DEBUG7 state_rms/head_rms ci #68: Commit ee10e71 pushed by rhCat
14s main
main
14s
main: DS4F_DEBUG7 -- the per-token top-5 logits + decoded tokens ci #67: Commit f107dad pushed by rhCat
16s main
main
16s
attn: fix the inverted F-rescale flip (was ON by default) ci #66: Commit 4bf90cb pushed by rhCat
18s main
main
18s
attn: F-rescale off by default (A/B: raw attention routes better) ci #65: Commit 9200de6 pushed by rhCat
11s main
main
11s
attn: DS4F_NO_F_RESCALE / DS4F_NO_STATE_RESCALE A/B toggles ci #64: Commit 2db9747 pushed by rhCat
19s main
main
19s
attn: M_PI -> literal (glibc strict C99 doesn't define it; Linux buil… ci #63: Commit f5263f0 pushed by rhCat
12s main
main
12s
attn: the tyrope -- yarn-style rope correction (V3.2/V4 rope_scaling) ci #62: Commit f0c2329 pushed by rhCat
10s main
main
10s
attn: SIMD row-range f8 matvec -- the threaded chunks decode with NEO… ci #61: Commit 1a18eeb pushed by rhCat
19s main
main
19s
attn: thread the big F8 matvecs (wo_a/wo_b/wq_b) across DS4F_ATTN_THR… ci #60: Commit 5a76eef pushed by rhCat
14s main
main
14s
attn: rope term v2 (standard rotary, theta = tyrope_theta) ci #59: Commit df59c7a pushed by rhCat
16s main
main
16s
attn: fix the two silent no-ops that kept attention off the real model ci #58: Commit edcffcf pushed by rhCat
19s main
main
19s
main: print the attention mode at startup ('real MLA' vs 'kvhalf fall… ci #57: Commit 00f8a9b pushed by rhCat
11s main
main
11s
attn: the real MLA (k3-structure, layout-driven) replaces the kvhalf … ci #56: Commit 3ffc141 pushed by rhCat
13s main
main
13s
mHC: state-rescale in the !up_ok fallback too + print the real up shape ci #55: Commit 23f8f0b pushed by rhCat
9s main
main
9s
main/Makefile: embed the git hash -- 'ds4f build <sha>' at startup ci #54: Commit 37b29e2 pushed by rhCat
13s main
main
13s
mHC: state-rescale to layer-input RMS (the F-rescale alone cannot bou… ci #53: Commit 7fd66a2 pushed by rhCat
9s main
main
9s
main: fail loudly on attn_step errors + print the hc tensor dtypes/sh… ci #52: Commit 9c045b8 pushed by rhCat
16s main
main
16s
mHC: F-rescale to layer-input RMS (bounded F; interim for the approxi… ci #51: Commit 968a951 pushed by rhCat
11s main
main
11s