so I have updated my repository of llama.cp: /home/eaman/llama/llama.cpp

And I should have applied the patch to the new pull before building,

problem is I think it did not work anymore:


with script /home/eaman/models/q6_roctest
bin directory for the latst build /home/eaman/llama/llama.cpp/build  is /home/eaman/llama/bin_vulkan_latest/

I get:

I common_memory_breakdown_print: | memory breakdown [MiB] | total    free     self   model   context   compute    unaccounted |
I common_memory_breakdown_print: |   - ROCm0 (RX 6800)    | 16368 = 15782 + (12493 = 11912 +     446 +     134) +      -11907 |
I common_memory_breakdown_print: |   - ROCm1 (RX 6700 XT) | 12272 = 11870 + ( 8906 =  8474 +     267 +     164) +       -8504 |
I common_memory_breakdown_print: |   - Host               |                   1022 =   994 +       0 +      28                |
I common_params_fit_impl: context size reduced from 262144 to 110080 -> need 6367 MiB less memory in total

So I run that same model without MTP: 


I common_memory_breakdown_print: | memory breakdown [MiB] | total    free     self   model   context   compute    unaccounted |
I common_memory_breakdown_print: |   - ROCm0 (RX 6800)    | 16368 = 16066 + (12212 = 11912 +     166 +     134) +      -11910 |
I common_memory_breakdown_print: |   - ROCm1 (RX 6700 XT) | 12272 = 12040 + ( 8738 =  8474 +      99 +     164) +       -8506 |
I common_memory_breakdown_print: |   - Host               |                   1022 =   994 +       0 +      28                |
I common_params_fit_impl: context size reduced from 262144 to 161024 -> need 4234 MiB less memory in total

This looks like too much vram is wasted, little ctx spare
----------------------------------------

So I ran the model with the backup of the origina llama release used to make the patch: /home/eaman/llama/bin_vulkan_correct/


I common_memory_breakdown_print: | memory breakdown [MiB] | total    free     self   model   context   compute    unaccounted |
I common_memory_breakdown_print: |   - ROCm0 (RX 6800)    | 16368 = 15784 + (12493 = 11912 +     446 +     134) +      -11909 |
I common_memory_breakdown_print: |   - ROCm1 (RX 6700 XT) | 12272 = 11872 + ( 8906 =  8474 +     267 +     164) +       -8506 |
I common_memory_breakdown_print: |   - Host               |                   1022 =   994 +       0 +      28                |
I common_params_fit_impl: context size reduced from 262144 to 139776 -> need 5123 MiB less memory in total


Now It looks to me as smething has changed with the latest pull of llama.cp, because with the old one I get 139776 ctx while the latest build gives me 110080 .

now can you check the build directory and see:

1st: if patch was applied correctly in /home/eaman/llama/llama.cpp/build
2nd: in case it was applied maybe something changed with the latest pull that invalidates the patch?
