[Bug]: Web profiling tab on DeepSeek V4 is 100 percent on "other"

[acedogblast]

Commit or version

v1.5.0

Environment

• Ubuntu 24.04 LTS
• AMD EPYC 7282 16-Core Processor
• 128GB DDR4 RAM

Reproduction steps

Run with web interface:
COLI_ALLOW_INSECURE_BIND=1 ./coli web --model [path/to/DeepSeek-V4-Flash] --host [some local IP] --port 8000

Run a simple small prompt.

Expected behavior

Statistics of performance should show up.

Actual behavior and logs

Only shows "No profiled turns yet — send a chat message and the breakdown appears here."

[acedogblast]

It seems that #882 would fix this.

[JustVugg]

You are right, and I confirmed the premise of #882 independently: deepseek_v4 emits none of the serve-protocol telemetry the dashboard parses. On the v1.5.0 binary, strings deepseek_v4 | grep -c EMAP is 0 — the engine never speaks HWINFO, TIERS, EMAP, PROF or HITS, so the Profiling tab has nothing to render and correctly says so.

The UI is fine, the server is fine, the engine is silent. Nothing on your side.

#882 is the fix and it is not mergeable yet — three CI checks are failing. I have asked @PwrBank to look at those. Keeping this open until it lands, and it will close with it.

Worth noting the same gap covers the Brain page and the hardware panel for DeepSeek V4, not just Profiling — so when #882 goes in you should get all three at once.

[acedogblast]

Tested again with the latest dev build (commit c2e1c3c).
Brain and Profiling tabs now show information however it seems that in the profiling tab the "Turn wall time by phase" is always 100 percent on the "other". All other metrics like I/O and matmul is always 0 percent. I don't think this is correct behavior. See screenshot below.

[JustVugg]

Fix is up — #1005. Diagnosis: the phase telemetry hardcoded everything but disk to zero, and the compute phase (matmul) was never measured — which is exactly what you'd see on a box with enough RAM, because the model sits warm in page cache so even disk_sec ≈ 0, and the whole turn folds into "other".

The fix adds a matmul timer on the expert store (same pattern GLM already uses for its phase graph), so V4 now reports I/O and matmul; the dashboard is unchanged, it just gets real numbers instead of zeros.

One ask, since you clearly have the hardware to run V4 with the Profiling tab live: once #1005 lands (or if you build the branch fix/v4-phase-telemetry), does the graph show matmul instead of 100% other? I verified it builds and stays token-exact here, but this dev box can't hold the 284B checkpoint right now to screenshot the live tab. Your confirmation closes this.

Note: attention time still folds into "other" for now — it's in a struct the store doesn't reach; a follow-up adds it the same way. But I/O + matmul is the bulk of the turn.

[acedogblast]

Tried with the new branch. Matmul now shows up.
See Screenshot below:

Another note: my server has 128GB of ram which is not enough to load the entire DeepSeekV4-Flash (about 156GB on Disk).

Output of free:
total used free shared buff/cache available
Mem: 131758336 113174860 3602456 11896 16113560 18583476
Swap: 0 0 0

[JustVugg]

Merged to dev (#1005). The Profiling tab's "Turn wall time by phase" now reports I/O and matmul for DeepSeek V4 instead of folding everything into "other" — the matmul phase (the real cost when the model sits warm in your RAM) was simply never measured before.

@acedogblast — since you have the box that runs V4 with the tab live: pull dev and check the graph shows matmul now. Build verified token-exact across all three platforms here; I just couldn't screenshot the live tab (this dev box is RAM-tight for the 284B checkpoint). Your confirmation closes this. One known gap, stated in the PR: attention time still folds into "other" for now — a follow-up adds it the same way, but I/O + matmul is the bulk of every turn.

[acedogblast]

Here is a screenshot with latest dev.