Spaces:

flozi00
/

traceforge

Running

App Files Files Community

traceforge

Commit History

Oct 29 6 days ago

18:04
space: blame the retry, not the first attempt, when a result repeats Running

flozi00

Verified signature
73a9a5a
17:58
space: compare tab, and say when the writer overrode the plan

flozi00

Verified signature
dba9b36
17:49
fix: triage labels are 1-based; splice applied them 0-based

flozi00

Verified signature
29e96a6
17:37
space: let the hash dropdowns accept a pasted hash

flozi00

Verified signature
d4588c7
17:33
space: compare tab — seed vs output diff with the reason for every edit

flozi00

Verified signature
c84bdd3
16:51
space: say plainly that no endpoint is configured

flozi00

Verified signature
dfe568e
16:47
drop the ZeroGPU local model: the Space always uses a remote endpoint

flozi00

Verified signature
501e011
16:39
revert: never put an HF token in a public Space

flozi00

Verified signature
57f5760
16:37
zero-gpu: bill GPU time to TRACEFORGE_HF_TOKEN when set

flozi00

Verified signature
4f04d60
16:33
space: dump the request object so we can see how auth arrives

flozi00

Verified signature
67f3161
16:30
space: echo the auth headers so we can see what ZeroGPU receives

flozi00

Verified signature
0f18a25
16:18
docs: xlarge is the full Blackwell, and the 16 tok/s was the MoE sync loop

flozi00

Verified signature
b6497e5
16:14
zero-gpu: size the token budget to the measured decode rate

flozi00

Verified signature
de2dd32
16:07
zero-gpu: keep the batched MoE off CPU, where the gather is 33x slower

flozi00

Verified signature
a186636
15:58
zero-gpu: prefetch the checkpoint at startup, not inside the GPU slice

flozi00

Verified signature
44e6ccf
15:48
zero-gpu: fall back to the expert loop if the batched path ever raises

flozi00

Verified signature
e408ca2
15:45
zero-gpu: batched MoE decode (no per-expert host syncs) + report container RAM

flozi00

Verified signature
34a5a51
15:33
zero-gpu: log the parameter device and VRAM so we can prove the weights are on the GPU

flozi00

Verified signature
cfca239
14:52
docs: describe the FP8 checkpoint, the MIG slice and the measured 16 tok/s

flozi00

Verified signature
2d55b77
14:43
zero-gpu: log the raw model head for triage debugging

flozi00

Verified signature
55b7487
14:35
zero-gpu: a function-style system prompt so the model labels instead of role-playing

flozi00

Verified signature
416dae0
14:28
zero-gpu: give triage the 1200 tokens the think block actually needs

flozi00

Verified signature
f6f144c
14:22
zero-gpu: always close the think block when thinking is off

flozi00

Verified signature
9fea984
14:16
zero-gpu: log whether the think block was skipped

flozi00

Verified signature
b7d7745
14:10
zero-gpu: skip the forced think block for triage labels

flozi00

Verified signature
890a260
14:02
zero-gpu: cap triage output at 600 tokens (4000 let it ramble past the slice)

flozi00

Verified signature
0e61936
13:52
zero-gpu: log per-call token counts and latency

flozi00

Verified signature
2667314
13:41
zero-gpu: show which GPU the local model actually runs on

flozi00

Verified signature
8ad5a2e
13:37
zero-gpu: build causal-conv1d for the LFM2 short-conv layers

flozi00

Verified signature
2da577f
13:29
zero-gpu: report the GPU and cut the window to fit a 120 s slice

flozi00

Verified signature
54df832
13:21
zero-gpu: size the rewrite window to the measured 14 tok/s slice budget

flozi00

Verified signature
a052ff5
13:20
zero-gpu: size the rewrite window to the measured 14 tok/s slice budget

flozi00

Verified signature
53367a3
13:18
zero-gpu: return token counts and tok/s from the local server

flozi00

Verified signature
07b9ced
13:15
zero-gpu: report tokens/s so the rewrite budget can be sized to the slice

flozi00

Verified signature
ddc9f01
13:00
zero-gpu: probe reports its own latency and allows for the think block

flozi00

Verified signature
86d2239
12:57
zero-gpu: dequantize block-wise FP8 MoE experts (128x128 scales)

flozi00

Verified signature
9199b01
12:52
zero-gpu: load the FP8 checkpoint (dequantized at load) instead of 48 GB bf16

flozi00

Verified signature
09e92f4
12:40
zero-gpu: an explicit TRACEFORGE_LOCAL_MODEL wins over the default mount

flozi00

Verified signature
64f9b85
12:38
zero-gpu: report volume read speed in local model status

flozi00

Verified signature
0580109
12:32
zero-gpu: run LFM2.5-24B-A2B on the xlarge slice from a mounted volume

flozi00

Verified signature
e2e8afb

Oct 27 8 days ago

16:08
traceforge space

flozi00

Verified signature
a8fbc2f
12:19
traceforge space

flozi00

Verified signature
2b7058b
12:09
traceforge space

flozi00

Verified signature
8150f46
11:59
traceforge space

flozi00

Verified signature
65052d9
11:50
traceforge space

flozi00

Verified signature
926b008
11:44
traceforge space

flozi00

Verified signature
c13b4f3
11:41
traceforge space

flozi00

Verified signature
15eb1c5
11:37
traceforge space

flozi00

Verified signature
dd2d347
11:34
traceforge space

flozi00

Verified signature
e2732c7
11:30
traceforge space

flozi00

Verified signature
c133e63

Commit History

Oct 29 6 days ago

space: blame the retry, not the first attempt, when a result repeats Running

space: compare tab, and say when the writer overrode the plan

fix: triage labels are 1-based; splice applied them 0-based

space: let the hash dropdowns accept a pasted hash

space: compare tab — seed vs output diff with the reason for every edit

space: say plainly that no endpoint is configured

drop the ZeroGPU local model: the Space always uses a remote endpoint

revert: never put an HF token in a public Space

zero-gpu: bill GPU time to TRACEFORGE_HF_TOKEN when set

space: dump the request object so we can see how auth arrives

space: echo the auth headers so we can see what ZeroGPU receives

docs: xlarge is the full Blackwell, and the 16 tok/s was the MoE sync loop

zero-gpu: size the token budget to the measured decode rate

zero-gpu: keep the batched MoE off CPU, where the gather is 33x slower

zero-gpu: prefetch the checkpoint at startup, not inside the GPU slice

zero-gpu: fall back to the expert loop if the batched path ever raises

zero-gpu: batched MoE decode (no per-expert host syncs) + report container RAM

zero-gpu: log the parameter device and VRAM so we can prove the weights are on the GPU

docs: describe the FP8 checkpoint, the MIG slice and the measured 16 tok/s

zero-gpu: log the raw model head for triage debugging

zero-gpu: a function-style system prompt so the model labels instead of role-playing

zero-gpu: give triage the 1200 tokens the think block actually needs

zero-gpu: always close the think block when thinking is off

zero-gpu: log whether the think block was skipped

zero-gpu: skip the forced think block for triage labels

zero-gpu: cap triage output at 600 tokens (4000 let it ramble past the slice)

zero-gpu: log per-call token counts and latency

zero-gpu: show which GPU the local model actually runs on

zero-gpu: build causal-conv1d for the LFM2 short-conv layers

zero-gpu: report the GPU and cut the window to fit a 120 s slice

zero-gpu: size the rewrite window to the measured 14 tok/s slice budget

zero-gpu: size the rewrite window to the measured 14 tok/s slice budget

zero-gpu: return token counts and tok/s from the local server

zero-gpu: report tokens/s so the rewrite budget can be sized to the slice

zero-gpu: probe reports its own latency and allows for the think block

zero-gpu: dequantize block-wise FP8 MoE experts (128x128 scales)

zero-gpu: load the FP8 checkpoint (dequantized at load) instead of 48 GB bf16

zero-gpu: an explicit TRACEFORGE_LOCAL_MODEL wins over the default mount

zero-gpu: report volume read speed in local model status

zero-gpu: run LFM2.5-24B-A2B on the xlarge slice from a mounted volume

Oct 27 8 days ago

traceforge space

traceforge space

traceforge space

traceforge space

traceforge space

traceforge space

traceforge space

traceforge space

traceforge space

traceforge space