#2309 prefill copy_i32: misaligned address

Descrição

My setup:

RTX 3080 Ti (Laptop 16GB - 175W)
64GB DDR5-4800
i9-12900HK

Model: **strata-iq3_s**

From the log file *strata-iq3_s.log*
<pre>
strata serve: route tail skip: 2341433 missed experts skipped (2414941 entries) since the start
strata serve: suffix drafts: 15 windows, 22 of 30 drafts accepted
strata serve: prompt 104282 tokens = 104117 reused + 165 read in 894 ms (184.5 tok/s), 114 generated in 2084 ms (54.7 tok/s), drafts accepted 59 of 84, 6 checkpoints
strata serve: decode expert cache hit rate: 70.5% (39564 hits / 56094 lookups); 1723 more read by the GPU over PCIe or from another GPU (3.0% of all 57817 routed)
strata serve: route tail skip: 2350248 missed experts skipped (2424116 entries) since the start
strata serve: suffix drafts: 7 windows, 16 of 20 drafts accepted
strata serve: prompt 105088 tokens = 104395 reused + 693 read in 1579 ms (438.8 tok/s), 265 generated in 4162 ms (63.7 tok/s), drafts accepted 161 of 178, 6 checkpoints
strata serve: decode expert cache hit rate: 73.5% (85402 hits / 116186 lookups); 3357 more read by the GPU over PCIe or from another GPU (2.8% of all 119543 routed)
strata serve: route tail skip: 2366545 missed experts skipped (2441138 entries) since the start
strata serve: prompt 107686 tokens = 105354 reused + 2332 read in 3985 ms (585.2 tok/s), 87 generated in 1266 ms (68.7 tok/s), drafts accepted 53 of 56, 6 checkpoints
strata serve: decode expert cache hit rate: 72.7% (26265 hits / 36139 lookups); 1129 more read by the GPU over PCIe or from another GPU (3.0% of all 37268 routed)
strata serve: route tail skip: 2372534 missed experts skipped (2447312 entries) since the start
strata serve: suffix drafts: 1 windows, 2 of 2 drafts accepted
prefill copy_i32: misaligned address
</pre>

Log from terminal
<pre>
[strata] answering: 1244 of max 81920 tokens, 52.1 tok/s, 95 s
[strata] writing a tool call: bash: 1306 of max 81920 tokens, 52.4 tok/s, 96 s
[strata] done: 1359 tokens in 97 s (52.6 tok/s) (stop, cancel=False), expert cache 71.1% hit (+2.6% of the routed experts over PCIe)
[strata] thinking: 2 of max 81920 tokens, 44.1 tok/s, 1 s
[strata] thinking: 48 of max 81920 tokens, 45.5 tok/s, 2 s
[strata] writing a tool call: bash: 112 of max 81920 tokens, 54.4 tok/s, 3 s
[strata] done: 114 tokens in 3 s (55.3 tok/s) (stop, cancel=False), expert cache 70.5% hit (+3.0% of the routed experts over PCIe)
[strata] reading the prompt: 105,083 of 105,088 tokens, 2 s so far
[strata] thinking: 48 of max 81920 tokens, 55.9 tok/s, 3 s
[strata] thinking: 123 of max 81920 tokens, 65.3 tok/s, 4 s
[strata] thinking: 195 of max 81920 tokens, 67.1 tok/s, 5 s
[strata] writing a tool call: read: 253 of max 81920 tokens, 64.2 tok/s, 6 s
[strata] done: 265 tokens in 6 s (64.0 tok/s) (stop, cancel=False), expert cache 73.5% hit (+2.8% of the routed experts over PCIe)
[strata] reading the prompt: 107,681 of 107,686 tokens, 4 s so far
[strata] writing a tool call: read: 48 of max 81920 tokens, 66.3 tok/s, 5 s
[strata] done: 87 tokens in 5 s (70.0 tok/s) (stop, cancel=False), expert cache 72.7% hit (+3.0% of the routed experts over PCIe)
[strata] the engine stopped unexpectedly (exit code 1). The engine exited (code 1). Its last log line: strata serve: suffix drafts: 1 windows, 2 of 2 drafts accepted - if that does not explain it, please report it at github.com/Niko1221/Strata/issues with the log. The next request starts the engine again. Its log: /home/dandandelion/AI/Inference/Strata/strata-iq3_s.log
[strata] done: 0 tokens in 4 s (0.0 tok/s) (error, cancel=False)
</pre>

I'm sorry I do not know if you'd need more info than what I have provided. If you want, you can point me to one and I'll get it for you. Also this seem to only happen on GSQ-RCO-IQ3_S and not in anything smaller like the IQ3_XXS or the Coder.