EAGLE4 drafter pair · box1 silicon lab
Tested Images — 2026-09-26
as of 19:53 EDT · covers 2026-09-24 → 2026-09-26
Every HEAD + TAIL image pair that ran on the two V80 cards, with its features, code, build and test scripts, and every sweep cell (bench × width × depth × unprimed / primed / KVI). Numbers are measured; tok/s and accept come from the lab's run history, rebuilt from the box1 arm logs at 19:53 EDT. Full text catalog: /shared/rnd_hq/2026-09-26/TESTED_IMAGES_20260926.md.
E1+CF E3+CF E7+F3 C7+F4 A8+F4 A5+F4 images GPU refs
Headline
mt d2 columns: unprimed, KVI off, FASTGLUE off, SGLang 0.5.14, 50×32, mean of every plain arm (n). Units: tok/s @ mean accept length.
Timeline
09-24 first silicon
Baseline pair (reference texts for every later pair)
HEAD E1 SEAM_WIDE = 1 KVONLY_SKIP_ATTN = 1
TAIL CF TREE_WIDTH_MAX = 8 TREE_WIDTH runtime reg 0x1CC
best 178.16 tok/s @ 1.83 — mt_bench w2 d2 unprimed, FASTGLUE off
09-25 first silicon
E1 + ATTN_PIPE
HEAD E3 ATTN_PIPE = 1 ATTN_CSA_PAR = 6
TAIL CF unchanged
best 180.74 tok/s @ 1.97 — mt_bench w3 d2 unprimed, FASTGLUE off
09-26 first silicon
256-lane HEAD + faster TAIL
HEAD E7 LANES_V = 256
TAIL F3 G12_OUTST3 = 1 S1_OVERLAP = 0 G1_LMDOWN_LANES_V = 64 PIPE_CUT_SILU_V = 0 parameter added, built at 0
best 190.75 tok/s @ 1.83 — mt_bench w2 d2 unprimed, FASTGLUE off
09-26 first silicon
KVI v2 HEAD (128 lanes) + G12X TAIL
HEAD C7 KVI_EN = 1 KVI_WIDE = 1 CHILD_ROPE_FIX = 1
TAIL F4 G12X_LANES_V = 64 TREE_P = 32
best 181.13 tok/s @ 1.98 — mt_bench w3 d2 unprimed, FASTGLUE off
09-26 first silicon
KVI v2 + 256 lanes; KVI, FASTGLUE and backend levers
HEAD A8 KVI_EN = 1 KVI_WIDE = 1 CHILD_ROPE_FIX = 1 LANES_V = 256 A8_TIMING_GUARD=1 build flow
TAIL F4 unchanged
best 233.73 tok/s @ 2.55 — mt_bench w2 d2 KVI, FASTGLUE v4
09-26 first silicon
Embedding fold + burst export: fold off (protocol, ADAPTIVE_TREE, KVI) then fold on
HEAD A5 G4_EMBFOLD = 1 G4_EMBFOLD_NPC = 9 SEAM_WIDE = 2
TAIL F4 unchanged
best 268.01 tok/s @ 2.55 — mt_bench w2 d2 KVI, FASTGLUE v4
Images
Features are cumulative; filled chips are what the image adds over the previously tested image of its line (F3 includes the never-flashed F1, A8 includes C5, A5 includes A3). Hover a chip for the exact strap text. WNS/WHS are post-route, ns.
md5 7b02c367 WNS +0.001 WHS +0.010
PIPE_ROLE = 2'd1 PERSIST = 1 DEPTH_N = 1 EMBG_EN = 1 FORREASON = 1 G8_SPLICE = 0 FR_KF_BANKS = 16 FR_ABANK_BRAM = 1 DUAL_PORT = 1 KV_RES_CLR_REP = 1 DBG_CAP_EN_V = 0 EAGLE4_OBS_LVL_P = 0 DERATE_MHZ 66 redesign_t0w2_kvonly_head source LANES_V = 128 HEAD_PAR = 8 ARRAY_FAST_CLK = 1 SEAM_WIDE = 1 KVONLY_SKIP_ATTN = 1
parent CD 0xE4B711CD
code asgard: /shared/build/gE1_headlever/verilog_43a8661/forks/head_seam256_kvskip_20260924 (EXPORT_MD5_LIST.txt, 317 files); box1: box1:~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/verilog/forks/head_seam256_kvskip_20260924; README: reports/results/HEAD_SEAM256_KVSKIP_20260924.md (repo)
commit 43a8661 (toggles added in 0220bf7; 43a8661 regenerates pkg_param_list.tcl)
build /home/y344shi/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/scripts/uni_pipeline_headlever_asgard.sh (md5 7fcd35cc; staged strap loop line 796)
wrapper /home/y344shi/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/scripts/run_headlever_0xE4B711E1_asgard.sh (plain; no vmpar2 variant existed; used per /shared/build/gE1_headlever/launch_asgard_4.log)
host asgard (mesh class pdi_build_asgard_e1, rc=0 2026-09-24 14:21)
host patch none new: E1/E3 already in oneflip_relay_driver KVONLY_HEAD_BUILDS (CONTEXT_DIGEST_20260924.md §8); LANES 128: no lane table entry needed
timing src source: /shared/build/gE1_headlever/aved_oneflip_p2s_head_asgard/hw/amd_v80_gen5x8_25.1_pdi23/build/prj.runs/impl_1/post_route_phys_opt_report_timing_summary_0.rpt; 0 failing of 2,585,701 endpoints per box1 ADJUDICATION_20260924.txt
PDI asgard ~/eagle4_artifacts/uni_oneflip/oneflip_p2s_head_asgard_0xE4B711E1_66m_7b02c367.pdi (clone: /shared/build/gE1_headlever/aved_oneflip_p2s_head_asgard/hw/amd_v80_gen5x8_25.1_pdi23/amd_v80_gen5x8_25.1.pdi); box1:~/eagle4_artifacts/uni_oneflip/same name
tested E1+CF
best 178.16 tok/s @ 1.83 on E1+CF (mt_bench w2 d2 unprimed, FASTGLUE off)
md5 3e3bc79a WNS +0.015 WHS +0.011
PIPE_ROLE = 2'd2 PERSIST = 1 DEPTH_N = 1 EMBG_EN = 1 FORREASON = 1 G8_SPLICE = 1 DERATE_MHZ 66 attention + RoPE pruned on TAIL TISEQ = 1 TREE_P = 16 SER_DEPTH = 16384 TILE_PIPE = 1 TREE_PP = 1 TREE_WIDTH_MAX = 8 TREE_WIDTH runtime reg 0x1CC
parent CB 0xE4B711CB
code asgard: /shared/build/tCF_width/verilog_ef36d00/forks/tree_width_tail_20260924 (staging for the withdrawn asgard copy; same md5 list); box1: box1:~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/verilog/forks/tree_width_tail_20260924; README: reports/results/TREE_WIDTH_PAIR_20260924.md (repo)
commit ef36d00
build /shared/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/scripts/uni_pipeline_tree_width_tail_olegvm_slot3.sh (md5 033ac7ea) (inferred from _hostwide_slot3 in GATE_VERDICTS + 'VM slot-3 copy was already routing')
wrapper —
host olegvm slot 3 (clone ~/alt_build_20260922/tCF_width; heavy lock _hostwide_slot3)
host patch PAIR_TREE_WIDTH host flag (scripts/pair_tree_width.py, arm_tail_width writes 0x1CC before each tree root; host commit 92f4e0be) for width < 8; width 8 needs nothing
timing src source: /shared/rnd_hq/hq_daily/images.csv; VM route report not harvested to asgard
PDI box1:~/eagle4_artifacts/uni_oneflip/oneflip_p2s_tail_olegvm_0xE4B711CF_66m_3e3bc79a.pdi (olegvm ~/eagle4_artifacts/uni_oneflip/ same name, not read)
tested E1+CF , E3+CF
best 180.74 tok/s @ 1.97 on E3+CF (mt_bench w3 d2 unprimed, FASTGLUE off)
md5 809219ea WNS +0.002 WHS +0.010
PIPE_ROLE = 2'd1 PERSIST = 1 DEPTH_N = 1 EMBG_EN = 1 FORREASON = 1 G8_SPLICE = 0 FR_KF_BANKS = 16 FR_ABANK_BRAM = 1 DUAL_PORT = 1 KV_RES_CLR_REP = 1 DBG_CAP_EN_V = 0 EAGLE4_OBS_LVL_P = 0 DERATE_MHZ 66 redesign_t0w2_kvonly_head source LANES_V = 128 HEAD_PAR = 8 ARRAY_FAST_CLK = 1 SEAM_WIDE = 1 KVONLY_SKIP_ATTN = 1 ATTN_PIPE = 1 ATTN_CSA_PAR = 6
parent E1 0xE4B711E1
code olegvm: ~/alt_build_20260922/gE3_attnpipe/verilog_ecddc77/forks/attn_pipe_ce_20260924 (EXPORT_MD5_LIST.txt, 320 files); box1: box1:~/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/verilog/forks/attn_pipe_ce_20260924; README: reports/results/LEVEL_BATCHED_HEAD_20260924.md §9-10 (repo)
commit ecddc77
build /shared/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/scripts/uni_pipeline_attnpipe_e3_olegvm.sh (md5 11ec767f; strap loop line 799, XCI want ATTN_CSA_PAR 6 line 1561)
wrapper /shared/workspace/eagle4_adaptation/hardware/cleanup_EAGLE/scripts/run_attnpipe_0xE4B711E3_olegvm.sh (plain; launched detached on the VM, not via mesh)
host olegvm slot 3 (clone ~/alt_build_20260922/gE3_attnpipe)
host patch none new: E3 already in KVONLY_HEAD_BUILDS (CONTEXT_DIGEST_20260924.md §8)
timing src source: box1:~/eagle4_artifacts/uni_oneflip/oneflip_p2s_head_olegvm_0xE4B711E3_66m_809219ea.ADJUDICATION_20260925.txt, citing VM route_report_timing_summary_0.rpt; 0 failing of 2,588,052
PDI box1:~/eagle4_artifacts/uni_oneflip/oneflip_p2s_head_olegvm_0xE4B711E3_66m_809219ea.pdi (VM source md5 identical)
tested E3+CF
best 180.74 tok/s @ 1.97 on E3+CF (mt_bench w3 d2 unprimed, FASTGLUE off)
md5 907f44ee WNS 0.000 WHS +0.008
PIPE_ROLE = 2'd1 PERSIST = 1 DEPTH_N = 1 EMBG_EN = 1 FORREASON = 1 G8_SPLICE = 0 FR_KF_BANKS = 16 FR_ABANK_BRAM = 1 DUAL_PORT = 1 KV_RES_CLR_REP = 1 DBG_CAP_EN_V = 0 EAGLE4_OBS_LVL_P = 0 DERATE_MHZ 66 redesign_t0w2_kvonly_head source LANES_V = 256 HEAD_PAR = 8 ARRAY_FAST_CLK = 1 SEAM_WIDE = 1 KVONLY_SKIP_ATTN = 1
parent E1 0xE4B711E1
code note: fork head_seam256_lanes256_20260925 (eagle4_config.svh md5 f554da2b; only diff vs E1 = line 326 LANES_V 128->256). Not found on asgard (/shared/scratch, /shared/build) or box1 verilog/forks; sim copy of the source: /shared/scratch/l256_head/src_L256; README: /shared/scratch/l256_head/handoff_host/README_E7_HOST.md; /shared/scratch/l256_head/L256_HEAD_FLASH_GATE_20260924.md
commit none (fork shipped as tarball/handoff dir, never committed; see code)
build scripts/uni_pipeline_headlever_e7_olegvm.sh (md5 b088bb70; E2 pipeline with SEAM_WIDE 2->1 in 2 places) - exists only on olegvm; not found on asgard or box1
wrapper olegvm ~/run_gE7_headlever256_seam1.sh (not on asgard/box1; plain, pre-vmpar2)
host olegvm
host patch L256 lane table: L256_HOST_PACKING.patch (LANES_BY_BUILD_ID {0xE4B711E7: 256} in eagle4_role_geometry.py + full_session_m7.py), later superseded by L256_HOST_PACKING_v3_E7_A6_A8_AA.patch; L256_RELAY_KVONLY.patch: add 0xE4B711E7 to KVONLY_HEAD_BUILDS (oneflip_relay_driver.py:2293, both $R and $W); box1 copies: /sanctuary/pdi_inbox/E7_HOST/
timing src source: /shared/rnd_hq/2026-09-25/PDI_INBOX_NOTE_E7_20260925.md; 0 failing endpoints
PDI box1:~/eagle4_artifacts/uni_oneflip/oneflip_p2s_head_olegvm_0xE4B711E7_66m_907f44ee.pdi; box1:/sanctuary/pdi_inbox/oneflip_p2s_head_olegvm_0xE4B711E7_66m_907f44ee.pdi
tested E7+F3
best 190.75 tok/s @ 1.83 on E7+F3 (mt_bench w2 d2 unprimed, FASTGLUE off)
md5 693a35da WNS +0.280 WHS +0.010
PIPE_ROLE = 2'd2 PERSIST = 1 DEPTH_N = 1 EMBG_EN = 1 FORREASON = 1 G8_SPLICE = 1 DERATE_MHZ 66 attention + RoPE pruned on TAIL TISEQ = 1 TREE_P = 16 SER_DEPTH = 16384 TILE_PIPE = 1 TREE_PP = 1 TREE_WIDTH_MAX = 8 TREE_WIDTH runtime reg 0x1CC G12_OUTST3 = 1 S1_OVERLAP = 0 G1_LMDOWN_LANES_V = 64 PIPE_CUT_SILU_V = 0
parent F1 0xE4B711F1 (never flashed; itself = CF 0xE4B711CF)
code asgard: /shared/scratch/tail_levers2/tail_levers2_fork_work_20260924.tar.gz (md5 a0f314bb) = /shared/scratch/tail_levers2/fork_work; staged /shared/build/tF3_g1lanes64/verilog_a0f314bb_forks_tail_levers2_g1_20260925; box1: not present; README: /shared/scratch/tail_levers2/TAIL_CHILD_LEVERS2_20260924.md §5 Build A; /shared/rnd_hq/2026-09-25/PDI_INBOX_NOTE_F1_F3_20260925.md
commit none (fork shipped as tarball/handoff dir, never committed; see code)
build scripts/uni_pipeline_tail_consolidated_g1_asgard_slot2.sh (md5 ae3fdd17; asgard copy /shared/scratch/tail_levers2/f4_ref/uni_pipeline_tail_consolidated_g1_asgard_slot2.sh)
wrapper /shared/scratch/tail_levers2/f4_ref/run_tF3_g1lanes64_olegvm2.sh (plain)
host olegvm2 (slot tF3_g1lanes64, lock _hostwide_slot2)
host patch none: host ABI unchanged vs CF (PDI_INBOX_NOTE_F1_F3)
timing src source: /shared/rnd_hq/2026-09-25/PDI_INBOX_NOTE_F1_F3_20260925.md; 0 failing endpoints
PDI box1:~/eagle4_artifacts/uni_oneflip/oneflip_p2s_tail_asgard_0xE4B711F3_66m_693a35da.pdi; box1:/sanctuary/pdi_inbox/same name
tested E7+F3
best 190.75 tok/s @ 1.83 on E7+F3 (mt_bench w2 d2 unprimed, FASTGLUE off)
md5 4e9014f6 WNS +0.004 WHS +0.010
PIPE_ROLE = 2'd1 PERSIST = 1 DEPTH_N = 1 EMBG_EN = 1 FORREASON = 1 G8_SPLICE = 0 FR_KF_BANKS = 16 FR_ABANK_BRAM = 1 DUAL_PORT = 1 KV_RES_CLR_REP = 1 DBG_CAP_EN_V = 0 EAGLE4_OBS_LVL_P = 0 DERATE_MHZ 66 redesign_t0w2_kvonly_head source LANES_V = 128 HEAD_PAR = 8 ARRAY_FAST_CLK = 1 SEAM_WIDE = 1 KVONLY_SKIP_ATTN = 1 ATTN_PIPE = 1 ATTN_CSA_PAR = 6 KVI_EN = 1 KVI_WIDE = 1 CHILD_ROPE_FIX = 1
parent E3 0xE4B711E3
code asgard: /shared/scratch/kvi/forks_v2/handoff_v2/head_kvi2_e3_20260925.tar.gz (md5 a9ffcd46); staged /shared/build/kvi2_stage_c7/forks/head_kvi2_e3_20260925; box1: not present; README: /shared/scratch/kvi/forks_v2/handoff_v2/README_HANDOFF_V2.md; /shared/rnd_hq/2026-09-25/PDI_INBOX_NOTE_C7_20260925.md
commit none (fork shipped as tarball/handoff dir, never committed; see code)
build /shared/build/kvi2_stage_c7/uni_pipeline_kvi2_e3.sh (md5 b4b5100c = handoff /shared/scratch/kvi/forks_v2/handoff_v2/uni_pipeline_kvi2_e3.sh)
wrapper /shared/build/kvi2_stage_c7/run_kvi2_0xE4B711C7.sh (plain, confirmed by /shared/build/kvi2_stage_c7/launch.log)
host asgard (mesh class pdi_build_head_kvi2_asgard, rc=0 2026-09-25 15:44, peak 72.5 GB)
host patch KVI host (KVI_HOST, flag EAGLE4_PAIR_KVI, HBM 0x42B0000000, CSRs 0x1E0..0x1F4); relay KVONLY build-id set: add 0xE4B711C7 to KVONLY_HEAD_BUILDS
timing src source: /shared/build/gC7_kvi2/aved_oneflip_p2s_head_asgard/hw/amd_v80_gen5x8_25.1_pdi23/build/prj.runs/impl_1/post_route_phys_opt_report_timing_summary_0.rpt
PDI asgard ~/eagle4_artifacts/uni_oneflip/oneflip_p2s_head_asgard_0xE4B711C7_66m_4e9014f6.pdi (clone /shared/build/gC7_kvi2/aved_oneflip_p2s_head_asgard/hw/amd_v80_gen5x8_25.1_pdi23/amd_v80_gen5x8_25.1.pdi); box1:~/eagle4_artifacts/uni_oneflip/same name; box1:/sanctuary/pdi_inbox/same name
tested C7+F4
best 181.13 tok/s @ 1.98 on C7+F4 (mt_bench w3 d2 unprimed, FASTGLUE off)
md5 85af2587 WNS +0.054 WHS +0.010
PIPE_ROLE = 2'd2 PERSIST = 1 DEPTH_N = 1 EMBG_EN = 1 FORREASON = 1 G8_SPLICE = 1 DERATE_MHZ 66 attention + RoPE pruned on TAIL TISEQ = 1 TREE_P = 32 SER_DEPTH = 16384 TILE_PIPE = 1 TREE_PP = 1 TREE_WIDTH_MAX = 8 TREE_WIDTH runtime reg 0x1CC G12_OUTST3 = 1 S1_OVERLAP = 0 G1_LMDOWN_LANES_V = 64 PIPE_CUT_SILU_V = 0 G12X_LANES_V = 64
parent F3 0xE4B711F3
code asgard: /shared/scratch/tail_levers2/handoff_f4/fork_tail_levers2_g12x_20260925 (+ .tar.gz md5 db185ccc); box1: not present; README: /shared/scratch/tail_levers2/handoff_f4/HANDOFF_F4.md
commit none (fork shipped as tarball/handoff dir, never committed; see code)
build /shared/scratch/tail_levers2/handoff_f4/uni_pipeline_tail_g12x_f4_vmpar2_asgard.sh (md5 35c75e4a, wrapper default) or uni_pipeline_tail_g12x_f4_asgard.sh (md5 4dfc4475)
wrapper /shared/scratch/tail_levers2/handoff_f4/run_tF4_kpar32_olegvm2.sh (uses _vmpar2 unless F4_PIPELINE overrides; which one actually ran: vmpar2 (inferred), no launch log on asgard)
host olegvm2 (images.csv; slot tF4_kpar32)
host patch none recorded in HANDOFF_F4.md (TAIL ABI unchanged) (inferred)
timing src source: /shared/rnd_hq/2026-09-25/PDI_INBOX_NOTE_BATCH_20260926.md; 0 failing endpoints
PDI box1:~/eagle4_artifacts/uni_oneflip/oneflip_p2s_tail_asgard_0xE4B711F4_66m_85af2587.pdi; box1:/sanctuary/pdi_inbox/same name
tested C7+F4 , A8+F4 , A5+F4
best 268.01 tok/s @ 2.55 on A5+F4 (mt_bench w2 d2 KVI, FASTGLUE v4)
md5 97e717ed WNS +0.002 WHS +0.010
PIPE_ROLE = 2'd1 PERSIST = 1 DEPTH_N = 1 EMBG_EN = 1 FORREASON = 1 G8_SPLICE = 0 FR_KF_BANKS = 16 FR_ABANK_BRAM = 1 DUAL_PORT = 1 KV_RES_CLR_REP = 1 DBG_CAP_EN_V = 0 EAGLE4_OBS_LVL_P = 0 DERATE_MHZ 66 redesign_t0w2_kvonly_head source LANES_V = 256 HEAD_PAR = 8 ARRAY_FAST_CLK = 1 SEAM_WIDE = 1 KVONLY_SKIP_ATTN = 1 KVI_EN = 1 KVI_WIDE = 1 CHILD_ROPE_FIX = 1 A8_TIMING_GUARD=1 build flow
parent C5 0xE4B711C5 (never flashed; itself = E1 0xE4B711E1)
code asgard: /shared/scratch/l256_head/handoff_a8/forks/head_kvi2_e1_l256_20260925 + MD5LIST_head_kvi2_e1_l256_20260925.txt; staged /shared/build/a8_stage/forks; box1: not present; README: /shared/scratch/l256_head/handoff_a8/README_A8.md
commit none (fork shipped as tarball/handoff dir, never committed; see code)
build /shared/build/a8_stage/uni_pipeline_a8_l256.sh (md5 0eb6e55e = /shared/scratch/l256_head/handoff_a8/uni_pipeline_a8_l256.sh)
wrapper /shared/build/a8_stage/run_0xE4B711A8.sh (plain, confirmed by /shared/build/a8_stage/launch.log)
host asgard (mesh class pdi_build_head_a8_l256, rc=0 2026-09-26 12:41)
host patch L256 lane table: L256_HOST_PACKING_v3_E7_A6_A8_AA.patch (160a1380); relay KVONLY build-id set: RELAY_KVONLY_ALL_v3.patch (+ .W for $W); KVI host: KVI_HOST_v3.patch (KVI_HEAD_BUILDS += A8); box1 copies: /sanctuary/pdi_inbox/L256_HOST/
timing src source: /shared/build/gA8_l256/aved_oneflip_p2s_head_asgard/hw/amd_v80_gen5x8_25.1_pdi23/build/prj.runs/impl_1/post_route_phys_opt_report_timing_summary_0.rpt
PDI asgard ~/eagle4_artifacts/uni_oneflip/oneflip_p2s_head_asgard_0xE4B711A8_66m_97e717ed.pdi (clone /shared/build/gA8_l256/...); box1:~/eagle4_artifacts/uni_oneflip/same name; box1:/sanctuary/pdi_inbox/same name
tested A8+F4
best 233.73 tok/s @ 2.55 on A8+F4 (mt_bench w2 d2 KVI, FASTGLUE v4)
md5 1f934249 WNS 0.000 WHS +0.001
PIPE_ROLE = 2'd1 PERSIST = 1 DEPTH_N = 1 EMBG_EN = 1 FORREASON = 1 G8_SPLICE = 0 FR_KF_BANKS = 16 FR_ABANK_BRAM = 1 DUAL_PORT = 1 KV_RES_CLR_REP = 1 DBG_CAP_EN_V = 0 EAGLE4_OBS_LVL_P = 0 DERATE_MHZ 66 redesign_t0w2_kvonly_head source LANES_V = 128 HEAD_PAR = 8 ARRAY_FAST_CLK = 1 SEAM_WIDE = 2 KVONLY_SKIP_ATTN = 1 ATTN_PIPE = 1 ATTN_CSA_PAR = 6 KVI_EN = 1 KVI_WIDE = 1 CHILD_ROPE_FIX = 1 G4_EMBFOLD = 1 G4_EMBFOLD_NPC = 9
parent A3 0xE4B711A3 (never flashed; itself = C7 0xE4B711C7)
code asgard: /shared/scratch/embfold_c7/handoff_a5/head_embfold_c7_20260925.tar.gz (same fork as A3); staged /shared/build/embfold_stage_a5/forks; box1: not present; README: /shared/scratch/embfold_c7/handoff_a5/README_HANDOFF_A5.md
commit none (fork shipped as tarball/handoff dir, never committed; see code)
build /shared/build/embfold_stage_a5/uni_pipeline_embfold_a5_vmpar2.sh (md5 bac03801 = /shared/scratch/embfold_c7/handoff_a5/uni_pipeline_embfold_a5_vmpar2.sh)
wrapper /shared/build/embfold_stage_a5/run_embfold_0xE4B711A5_vmpar2.sh (the _vmpar2 wrapper was used even though the host was asgard; confirmed by /shared/build/embfold_stage_a5/launch.log)
host asgard (mesh class pdi_build_head_embfold_a5, rc=0 2026-09-25 20:51)
host patch same as A3: EMBFOLD table upload + relay hook + .embfold relay/kvi_host (A5 in KVONLY_HEAD_BUILDS / KVI_HEAD_BUILDS); KVI host as C7; SEAM_WIDE 2 burst: host relay unchanged (per A2 handoff, PDI_INBOX_NOTE_BATCH)
timing src source: /shared/build/gA5_stack/aved_oneflip_p2s_head_asgard/hw/amd_v80_gen5x8_25.1_pdi23/build/prj.runs/impl_1/post_route_phys_opt_report_timing_summary_0.rpt
PDI asgard ~/eagle4_artifacts/uni_oneflip/oneflip_p2s_head_asgard_0xE4B711A5_66m_1f934249.pdi (clone /shared/build/gA5_stack/...); box1:~/eagle4_artifacts/uni_oneflip/same name; box1:/sanctuary/pdi_inbox/EMBFOLD/same name
tested A5+F4
best 268.01 tok/s @ 2.55 on A5+F4 (mt_bench w2 d2 KVI, FASTGLUE v4)
Pairs: test runs and sweeps
E1+CF HEAD E1 0xE4B711E1 + TAIL CF 0xE4B711CF
Baseline pair (reference texts for every later pair)
tok/s vs depth (50×32; hover a point for accept, rounds and source)
0
50
100
150
200
2
3
4
5
6
7
8
10
15
tree depth d
tok/s
hum unprimed w3 d2: 119.98 tok/s @ 2.06 accept (n=3; width_ab_20260924_230441.txt width_ab_20260924_231612.txt)
hum unprimed w3 d3: 105.12 tok/s @ 2.22 accept (n=1; width_ab_20260924_231612.txt)
hum unprimed w3 d5: 85.52 tok/s @ 2.38 accept (n=1; width_ab_20260924_231612.txt)
hum unprimed w3 d8: 64.55 tok/s @ 2.40 accept (n=1; width_ab_20260924_231612.txt)
hum unprimed w3 d10: 55.05 tok/s @ 2.40 accept (n=1; width_ab_20260924_231612.txt)
hum unprimed w3 d15: 40.20 tok/s @ 2.38 accept (n=1; width_ab_20260924_231612.txt)
hum primed w3 d2: 78.65 tok/s @ 2.88 accept (n=1; width_ab_20260924_232332.txt)
hum primed w3 d3: 78.20 tok/s @ 3.82 accept (n=1; width_ab_20260924_232332.txt)
hum primed w3 d5: 72.89 tok/s @ 5.08 accept (n=1; width_ab_20260924_232332.txt)
hum primed w3 d8: 69.71 tok/s @ 6.67 accept (n=1; width_ab_20260924_232332.txt)
hum primed w3 d10: 67.84 tok/s @ 7.44 accept (n=1; width_ab_20260924_232332.txt)
hum primed w3 d15: 61.13 tok/s @ 8.60 accept (n=1; width_ab_20260924_232332.txt)
mt unprimed w3 d2: 175.90 tok/s @ 1.97 accept (n=3; width_ab_20260924_223134.txt)
mt unprimed w3 d3: 142.22 tok/s @ 2.08 accept (n=1; width_ab_20260924_224558.txt)
mt unprimed w3 d5: 102.85 tok/s @ 2.14 accept (n=1; width_ab_20260924_224558.txt)
mt unprimed w3 d8: 71.33 tok/s @ 2.15 accept (n=1; width_ab_20260924_224558.txt)
mt unprimed w3 d10: 58.56 tok/s @ 2.14 accept (n=1; width_ab_20260924_224558.txt)
mt unprimed w3 d15: 41.39 tok/s @ 2.15 accept (n=1; width_ab_20260924_224558.txt)
mt primed w3 d2: 118.57 tok/s @ 2.61 accept (n=3; width_ab_20260924_224150.txt width_ab_20260924_235301.txt)
mt primed w3 d3: 110.52 tok/s @ 3.15 accept (n=3; width_ab_20260924_224150.txt width_ab_20260924_225205.txt width_ab_20260924_235301.txt)
mt primed w3 d4: 104.75 tok/s @ 3.46 accept (n=1; width_ab_20260924_235301.txt)
mt primed w3 d5: 97.00 tok/s @ 3.79 accept (n=2; width_ab_20260924_225205.txt width_ab_20260924_235301.txt)
mt primed w3 d6: 91.27 tok/s @ 4.00 accept (n=1; width_ab_20260924_235301.txt)
mt primed w3 d7: 86.28 tok/s @ 4.21 accept (n=2; width_ab_20260924_231439.txt width_ab_20260924_235301.txt)
mt primed w3 d8: 83.65 tok/s @ 4.36 accept (n=1; width_ab_20260924_225205.txt)
mt primed w3 d10: 73.78 tok/s @ 4.52 accept (n=1; width_ab_20260924_225205.txt)
mt primed w3 d15: 57.70 tok/s @ 4.68 accept (n=2; width_ab_20260924_225205.txt width_ab_20260924_234352.txt)
■ hum primed w3
■ mt primed w3
■ mt unprimed w3
■ hum unprimed w3
w × d grid — humaneval unprimed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32
source: width_ab_20260924_230441.txt, width_ab_20260924_231612.txt
Depth ladders and width rows
w3 · humaneval primed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32
d2 78.65 @ 2.88 · n=1 d3 78.20 @ 3.82 · n=1 d5 72.89 @ 5.08 · n=1 d8 69.71 @ 6.67 · n=1 d10 67.84 @ 7.44 · n=1 d15 61.13 @ 8.60 · n=1
source: width_ab_20260924_232332.txt
w3 · mt_bench primed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32
d2 118.57 @ 2.61 · n=3 d3 110.52 @ 3.15 · n=3 d4 104.75 @ 3.46 · n=1 d5 97.00 @ 3.79 · n=2 d6 91.27 @ 4.00 · n=1 d7 86.28 @ 4.21 · n=2 d8 83.65 @ 4.36 · n=1 d10 73.78 @ 4.52 · n=1 d15 57.70 @ 4.68 · n=2
source: width_ab_20260924_224150.txt, width_ab_20260924_225205.txt, width_ab_20260924_231439.txt, width_ab_20260924_234352.txt, width_ab_20260924_235301.txt
w3 · mt_bench primed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32, flags: TREE_PRIME_ROWS=16
d7 103.23 @ 3.05 · n=1 d15 57.47 @ 3.16 · n=1
source: width_ab_20260924_231229.txt, width_ab_20260924_234003.txt
w3 · mt_bench primed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32, flags: TREE_PRIME_ROWS=64
d7 108.67 @ 3.90 · n=1 d15 67.90 @ 4.27 · n=1
source: width_ab_20260924_231335.txt, width_ab_20260924_234122.txt
w3 · mt_bench unprimed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32
d3 142.22 @ 2.08 · n=1 d5 102.85 @ 2.14 · n=1 d8 71.33 @ 2.15 · n=1 d10 58.56 @ 2.14 · n=1 d15 41.39 @ 2.15 · n=1
source: width_ab_20260924_224558.txt
d2 · mt_bench unprimed, FASTGLUE off, SGLang 0.5.14, fa4, 50x32
w2 178.16 @ 1.83 · n=3 w3 175.90 @ 1.97 · n=3 w8 140.50 @ 2.21 · n=3
All 41 cells (best-by-cell, with rounds and source)
Census, 1 request (per-phase cycles)
E3+CF HEAD E3 0xE4B711E3 + TAIL CF 0xE4B711CF
E1 + ATTN_PIPE
tok/s vs depth (50×32; hover a point for accept, rounds and source)
0
50
100
150
2
3
5
8
10
tree depth d
tok/s
hum unprimed w3 d2: 120.16 tok/s @ 2.06 accept (n=1; width_ab_20260925_023719.txt)
hum unprimed w3 d3: 103.96 tok/s @ 2.22 accept (n=1; width_ab_20260925_023719.txt)
hum unprimed w3 d5: 84.43 tok/s @ 2.38 accept (n=1; width_ab_20260925_023719.txt)
hum unprimed w3 d8: 66.51 tok/s @ 2.40 accept (n=1; width_ab_20260925_023719.txt)
hum unprimed w3 d10: 56.13 tok/s @ 2.40 accept (n=1; arm log only)
■ hum unprimed w3
Depth ladders and width rows
w3 · humaneval unprimed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32
d2 120.16 @ 2.06 · n=1 d3 103.96 @ 2.22 · n=1 d5 84.43 @ 2.38 · n=1 d8 66.51 @ 2.40 · n=1 d10 56.13 @ 2.40 · n=1
source: arm, log, only, width_ab_20260925_023719.txt
w3 · mt_bench primed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32
d2 116.72 @ 2.61 · n=2 d3 108.36 @ 3.15 · n=1
source: width_ab_20260925_021936.txt, width_ab_20260925_023014.txt
d2 · mt_bench unprimed, FASTGLUE off, SGLang 0.5.14, fa4, 50x32
w2 175.39 @ 1.83 · n=4 w3 180.74 @ 1.97 · n=3
All 13 cells (best-by-cell, with rounds and source)
Census, 1 request (per-phase cycles)
E7+F3 HEAD E7 0xE4B711E7 + TAIL F3 0xE4B711F3
256-lane HEAD + faster TAIL
tok/s vs depth (50×32; hover a point for accept, rounds and source)
0
50
100
150
2
3
5
8
10
15
tree depth d
tok/s
hum unprimed w3 d2: 125.18 tok/s @ 2.06 accept (n=1; width_ab_20260926_120616.txt)
hum unprimed w3 d3: 112.38 tok/s @ 2.22 accept (n=1; width_ab_20260926_120616.txt)
hum unprimed w3 d5: 93.37 tok/s @ 2.38 accept (n=1; width_ab_20260926_120616.txt)
hum unprimed w3 d8: 72.85 tok/s @ 2.40 accept (n=1; width_ab_20260926_120616.txt)
hum unprimed w3 d10: 62.92 tok/s @ 2.40 accept (n=1; width_ab_20260926_120616.txt)
hum unprimed w3 d15: 46.96 tok/s @ 2.38 accept (n=1; width_ab_20260926_120616.txt)
hum primed w3 d2: 84.31 tok/s @ 2.88 accept (n=1; width_ab_20260926_121324.txt)
hum primed w3 d3: 82.58 tok/s @ 3.82 accept (n=1; width_ab_20260926_121324.txt)
hum primed w3 d5: 77.75 tok/s @ 5.08 accept (n=1; width_ab_20260926_121324.txt)
hum primed w3 d8: 73.06 tok/s @ 6.67 accept (n=1; width_ab_20260926_121324.txt)
hum primed w3 d10: 71.53 tok/s @ 7.44 accept (n=1; width_ab_20260926_121324.txt)
hum primed w3 d15: 65.32 tok/s @ 8.60 accept (n=1; width_ab_20260926_121324.txt)
■ hum primed w3
■ hum unprimed w3
Depth ladders and width rows
w3 · humaneval primed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32
d2 84.31 @ 2.88 · n=1 d3 82.58 @ 3.82 · n=1 d5 77.75 @ 5.08 · n=1 d8 73.06 @ 6.67 · n=1 d10 71.53 @ 7.44 · n=1 d15 65.32 @ 8.60 · n=1
source: width_ab_20260926_121324.txt
w3 · humaneval unprimed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32
d2 125.18 @ 2.06 · n=1 d3 112.38 @ 2.22 · n=1 d5 93.37 @ 2.38 · n=1 d8 72.85 @ 2.40 · n=1 d10 62.92 @ 2.40 · n=1 d15 46.96 @ 2.38 · n=1
source: width_ab_20260926_120616.txt
d2 · mt_bench unprimed, FASTGLUE off, SGLang 0.5.14, fa4, 50x32
w2 190.75 @ 1.83 · n=4 w3 188.50 @ 1.97 · n=3
All 18 cells (best-by-cell, with rounds and source)
Census, 1 request (per-phase cycles)
C7+F4 HEAD C7 0xE4B711C7 + TAIL F4 0xE4B711F4
KVI v2 HEAD (128 lanes) + G12X TAIL
tok/s vs depth (50×32; hover a point for accept, rounds and source)
0
50
100
150
2
3
5
8
10
tree depth d
tok/s
hum unprimed w3 d2: 119.27 tok/s @ 2.07 accept (n=1; width_ab_20260926_141459.txt)
hum unprimed w3 d3: 103.75 tok/s @ 2.24 accept (n=1; width_ab_20260926_141459.txt)
hum unprimed w3 d5: 86.67 tok/s @ 2.39 accept (n=1; width_ab_20260926_141459.txt)
hum unprimed w3 d8: 67.26 tok/s @ 2.41 accept (n=1; width_ab_20260926_141459.txt)
hum unprimed w3 d10: 57.42 tok/s @ 2.40 accept (n=1; arm log only)
■ hum unprimed w3
Depth ladders and width rows
w3 · humaneval unprimed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32
d2 119.27 @ 2.07 · n=1 d3 103.75 @ 2.24 · n=1 d5 86.67 @ 2.39 · n=1 d8 67.26 @ 2.41 · n=1 d10 57.42 @ 2.40 · n=1
source: arm, log, only, width_ab_20260926_141459.txt
d2 · mt_bench unprimed, FASTGLUE off, SGLang 0.5.14, fa4, 50x32
w2 177.84 @ 1.83 · n=3 w3 181.13 @ 1.98 · n=3
All 11 cells (best-by-cell, with rounds and source)
Census, 1 request (per-phase cycles)
A8+F4 HEAD A8 0xE4B711A8 + TAIL F4 0xE4B711F4
KVI v2 + 256 lanes; KVI, FASTGLUE and backend levers
tok/s vs depth (50×32; hover a point for accept, rounds and source)
0
50
100
150
200
250
2
3
4
5
7
10
15
tree depth d
tok/s
hum primed w3 d2: 82.81 tok/s @ 2.88 accept (n=1; width_ab_20260926_164034.txt)
hum primed w3 d5: 77.21 tok/s @ 5.10 accept (n=1; width_ab_20260926_164034.txt)
hum primed w3 d10: 73.40 tok/s @ 7.51 accept (n=1; width_ab_20260926_164034.txt)
hum primed w3 d15: 66.49 tok/s @ 8.74 accept (n=1; width_ab_20260926_164034.txt)
mt unprimed w3 d2: 187.74 tok/s @ 1.98 accept (n=1; width_ab_20260926_163152.txt)
mt unprimed w3 d3: 155.29 tok/s @ 2.09 accept (n=1; width_ab_20260926_163152.txt)
mt unprimed w3 d5: 118.80 tok/s @ 2.15 accept (n=1; width_ab_20260926_163152.txt)
mt unprimed w3 d7: 92.60 tok/s @ 2.15 accept (n=1; width_ab_20260926_163152.txt)
mt primed w3 d2: 126.72 tok/s @ 2.62 accept (n=1; width_ab_20260926_162727.txt)
mt primed w3 d3: 119.18 tok/s @ 3.19 accept (n=1; width_ab_20260926_162727.txt)
mt primed w3 d5: 109.00 tok/s @ 3.80 accept (n=1; width_ab_20260926_162727.txt)
mt primed w3 d7: 96.98 tok/s @ 4.22 accept (n=1; width_ab_20260926_162727.txt)
mt KVI w3 d2: 217.18 tok/s @ 2.61 accept (n=17; width_ab_20260926_162324.txt width_ab_20260926_164605.txt width_ab_20260926_164803.txt width_ab_20260926_165133.txt width_ab_20260926_165502.txt width_ab_20260926_165830.txt width_ab_20260926_170029.txt width_ab_20260926_170228.txt width_ab_20260926_170544.txt width_ab_20260926_170839.txt width_ab_20260926_171417.txt width_ab_20260926_172344.txt width_ab_20260926_172730.txt width_ab_20260926_173323.txt width_ab_20260926_180328.txt width_ab_20260926_181252.txt width_ab_20260926_181638.txt)
mt KVI w3 d3: 201.25 tok/s @ 3.20 accept (n=7; width_ab_20260926_162324.txt width_ab_20260926_164605.txt width_ab_20260926_164803.txt width_ab_20260926_165133.txt width_ab_20260926_165502.txt width_ab_20260926_165830.txt width_ab_20260926_170029.txt)
mt KVI w3 d5: 173.87 tok/s @ 3.88 accept (n=1; width_ab_20260926_162324.txt)
mt KVI w3 d7: 149.91 tok/s @ 4.35 accept (n=1; width_ab_20260926_162324.txt)
hum KVI w3 d2: 137.64 tok/s @ 2.89 accept (n=1; width_ab_20260926_163609.txt)
hum KVI w3 d5: 128.20 tok/s @ 5.14 accept (n=1; width_ab_20260926_163609.txt)
hum KVI w3 d10: 116.45 tok/s @ 7.62 accept (n=1; width_ab_20260926_163609.txt)
hum KVI w3 d15: 101.64 tok/s @ 8.99 accept (n=1; width_ab_20260926_163609.txt)
mt KVI+FASTGLUE4 w2 d2: 233.73 tok/s @ 2.55 accept (n=7; width_ab_20260926_171610.txt width_ab_20260926_171957.txt width_ab_20260926_172923.txt width_ab_20260926_173719.txt width_ab_20260926_180521.txt width_ab_20260926_180907.txt width_ab_20260926_181829.txt)
mt KVI+FASTGLUE4 w2 d3: 222.27 tok/s @ 3.01 accept (n=1; width_ab_20260926_173719.txt)
mt KVI+FASTGLUE4 w2 d4: 205.73 tok/s @ 3.31 accept (n=1; width_ab_20260926_173719.txt)
mt KVI+FASTGLUE4 w2 d5: 195.34 tok/s @ 3.64 accept (n=1; width_ab_20260926_173719.txt)
mt KVI+FASTGLUE4 w2 d7: 168.02 tok/s @ 3.98 accept (n=1; width_ab_20260926_173719.txt)
■ mt KVI+FASTGLUE4 w2
■ mt KVI w3
■ hum KVI w3
■ mt primed w3
■ mt unprimed w3
■ hum primed w3
w × d grid — mt_bench KVI, FASTGLUE off, SGLang 0.5.16, fa4, 50x32, flags: PAIR_SKIP_GPU_EXTEND=0 TREE_OVERLAP_CATCHUP=0
source: width_ab_20260926_162324.txt, width_ab_20260926_164605.txt, width_ab_20260926_164803.txt, width_ab_20260926_165133.txt, width_ab_20260926_165502.txt, width_ab_20260926_165830.txt, width_ab_20260926_170029.txt, width_ab_20260926_170228.txt, width_ab_20260926_170544.txt, width_ab_20260926_170839.txt, width_ab_20260926_171417.txt, width_ab_20260926_172344.txt, width_ab_20260926_172730.txt, width_ab_20260926_173323.txt, width_ab_20260926_180328.txt, width_ab_20260926_181252.txt, width_ab_20260926_181638.txt
w × d grid — mt_bench KVI, FASTGLUE v4, SGLang 0.5.16, fa4, 50x32, flags: PAIR_SKIP_GPU_EXTEND=0
source: width_ab_20260926_170643.txt, width_ab_20260926_170938.txt, width_ab_20260926_171610.txt, width_ab_20260926_171957.txt, width_ab_20260926_172923.txt, width_ab_20260926_173422.txt, width_ab_20260926_173719.txt, width_ab_20260926_180521.txt, width_ab_20260926_180907.txt, width_ab_20260926_181829.txt
Depth ladders and width rows
w3 · humaneval KVI, FASTGLUE off, SGLang 0.5.16, fa4, 50x32, flags: PAIR_SKIP_GPU_EXTEND=0 TREE_OVERLAP_CATCHUP=0
d2 137.64 @ 2.89 · n=1 d5 128.20 @ 5.14 · n=1 d10 116.45 @ 7.62 · n=1 d15 101.64 @ 8.99 · n=1
source: width_ab_20260926_163609.txt
w3 · humaneval primed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32
d2 82.81 @ 2.88 · n=1 d5 77.21 @ 5.10 · n=1 d10 73.40 @ 7.51 · n=1 d15 66.49 @ 8.74 · n=1
source: width_ab_20260926_164034.txt
w3 · mt_bench KVI, FASTGLUE off, SGLang 0.5.16, fa3, 50x32, flags: PAIR_SKIP_GPU_EXTEND=0 TREE_OVERLAP_CATCHUP=0
d2 218.92 @ 2.61 · n=3 d3 203.26 @ 3.20 · n=3
source: width_ab_20260926_165002.txt, width_ab_20260926_165331.txt, width_ab_20260926_165700.txt
w3 · mt_bench primed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32
d2 126.72 @ 2.62 · n=1 d3 119.18 @ 3.19 · n=1 d5 109.00 @ 3.80 · n=1 d7 96.98 @ 4.22 · n=1
source: width_ab_20260926_162727.txt
w3 · mt_bench unprimed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32
d2 187.74 @ 1.98 · n=1 d3 155.29 @ 2.09 · n=1 d5 118.80 @ 2.15 · n=1 d7 92.60 @ 2.15 · n=1
source: width_ab_20260926_163152.txt
d2 · mt_bench KVI, FASTGLUE v4, SGLang 0.5.16, fa4, 50x32, flags: PAIR_KVI_OVERLAP=1 PAIR_SKIP_GPU_EXTEND=0
w2 231.72 @ 2.55 · n=6 w3 219.80 @ 2.61 · n=7
d2 · mt_bench unprimed, FASTGLUE off, SGLang 0.5.14, fa4, 50x32
w2 188.91 @ 1.83 · n=9 w3 187.87 @ 1.98 · n=7
d2 · mt_bench unprimed, FASTGLUE v1, SGLang 0.5.14, fa4, 50x32
w2 175.20 @ 1.83 · n=4 w3 186.39 @ 1.98 · n=3
d2 · mt_bench unprimed, FASTGLUE v2, SGLang 0.5.14, fa4, 50x32
w2 187.13 @ 1.83 · n=8 w3 188.58 @ 1.98 · n=6
d2 · mt_bench unprimed, FASTGLUE v3, SGLang 0.5.14, fa4, 50x32
w2 204.50 @ 1.83 · n=4 w3 202.37 @ 1.98 · n=3
All 65 cells (best-by-cell, with rounds and source)
Census, 1 request (per-phase cycles)
A5+F4 HEAD A5 0xE4B711A5 + TAIL F4 0xE4B711F4
Embedding fold + burst export: fold off (protocol, ADAPTIVE_TREE, KVI) then fold on
tok/s vs depth (50×32; hover a point for accept, rounds and source)
0
50
100
150
200
250
300
2
3
4
5
10
15
tree depth d
tok/s
mt unprimed w3 d2: 201.36 tok/s @ 1.98 accept (n=1; width_ab_20260926_191420.txt)
mt unprimed w3 d3: 172.70 tok/s @ 2.09 accept (n=1; width_ab_20260926_191420.txt)
mt unprimed w3 d4: 151.38 tok/s @ 2.12 accept (n=1; width_ab_20260926_191420.txt)
hum KVI w3 d2: 153.21 tok/s @ 2.89 accept (n=1; width_ab_20260926_193355.txt)
hum KVI w3 d5: 153.03 tok/s @ 5.14 accept (n=1; width_ab_20260926_193355.txt)
hum KVI w3 d10: 144.67 tok/s @ 7.62 accept (n=1; width_ab_20260926_193355.txt)
hum KVI w3 d15: 132.19 tok/s @ 8.99 accept (n=1; width_ab_20260926_193355.txt)
mt KVI+FASTGLUE4 w1 d2: 259.45 tok/s @ 2.32 accept (n=1; width_ab_20260926_185929.txt)
mt KVI+FASTGLUE4 w1 d3: 259.20 tok/s @ 2.66 accept (n=1; width_ab_20260926_185929.txt)
mt KVI+FASTGLUE4 w1 d4: 252.39 tok/s @ 2.87 accept (n=1; width_ab_20260926_185929.txt)
mt KVI+FASTGLUE4 w1 d5: 244.10 tok/s @ 3.07 accept (n=1; width_ab_20260926_185929.txt)
■ mt KVI+FASTGLUE4 w1
■ mt unprimed w3
■ hum KVI w3
w × d grid — mt_bench KVI, FASTGLUE v4, SGLang 0.5.16, fa4, 50x32, flags: PAIR_SKIP_GPU_EXTEND=0
source: width_ab_20260926_185929.txt, width_ab_20260926_193006.txt
w × d grid — mt_bench unprimed, FASTGLUE off, SGLang 0.5.16, fa4, 50x32
source: arm, log, only, width_ab_20260926_191420.txt
Depth ladders and width rows
w3 · humaneval KVI, FASTGLUE v4, SGLang 0.5.16, fa4, 50x32, flags: PAIR_SKIP_GPU_EXTEND=0
d2 153.21 @ 2.89 · n=1 d5 153.03 @ 5.14 · n=1 d10 144.67 @ 7.62 · n=1 d15 132.19 @ 8.99 · n=1
source: width_ab_20260926_193355.txt
d7 · mt_bench KVI, FASTGLUE v4, SGLang 0.5.16, fa4, 50x32, flags: PAIR_SCORE_TRACE=1 PAIR_SKIP_GPU_EXTEND=0
w2 214.76 @ 3.98 · n=1 w3 197.03 @ 4.35 · n=1
d2 · mt_bench unprimed, FASTGLUE off, SGLang 0.5.14, fa4, 50x32
w2 202.51 @ 1.83 · n=4 w3 201.81 @ 1.98 · n=4
d2 · mt_bench unprimed, FASTGLUE off, SGLang 0.5.14, fa4, 50x32, flags: EMBFOLD_TABLE=/sanctuary/pdi_inbox/EMBFOLD/table_osprey_npc9_c7 EMBFOLD_WEIGHTS=/home/y344shi/.claude/jobs/5b98fbdb/tmp/fold_lineage
w2 203.79 @ 1.83 · n=4 w3 203.92 @ 1.98 · n=4
All 41 cells (best-by-cell, with rounds and source)
Census, 1 request (per-phase cycles)
GPU-only references
SGLang EAGLE on the H200 without the FPGA pair. Rows with a 345 MHz clock had the GPU clock sampled before warm-up.
Sources: regenerated lab history in /shared/rnd_hq/2026-09-26/tested_images_data/; session reports under /shared/rnd_hq/from_box1/; image metadata from hq_daily/images.csv and the handoff READMEs named on each card. Per-image CSV: /shared/rnd_hq/hq_daily/tested_images.csv.