Reduce small model context and rebalance GPU layers

docs/LLM/dealign-ornith-crack.md

@@ -45,19 +45,21 @@ The non-expert tensors total 2.02 GiB. The layer totals are 660.96 MiB for
 ordinary SSM blocks, 655.81 MiB for full-attention blocks, and 662.40 MiB for
 block 40. The expert portion is 630 MiB in every block.
 
-The service keeps all non-expert tensors on CUDA. It also keeps expert tensors
-from blocks 36 through 40 on CUDA. Those blocks sit next to the output and
-include the bundled MTP head. Expert tensors from blocks 0 through 35 move to
-CPU through this explicit override:
+The provisional service keeps all non-expert tensors on CUDA. It also keeps
+expert tensors from blocks 33 through 40 on CUDA. Those blocks sit next to the
+output and include the bundled MTP head. Expert tensors from blocks 0 through
+32 move to CPU through this explicit override:
 
 ```
-OVERRIDE_EXPERT='blk\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32|33|34|35)\.ffn_(up|down|gate|gate_up)_(ch|)exps'
+OVERRIDE_EXPERT='blk\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32)\.ffn_(up|down|gate|gate_up)_(ch|)exps'
 OVERRIDE_BUFT=CPU
 ```
 
-This places 3.69 GiB of expert weights on CUDA and 22.13 GiB on CPU. The
-selection leaves the output-adjacent blocks and MTP path on CUDA while the
-f16 context cache uses the remaining GPU budget.
+The previous five-block configuration placed 3.69 GiB of expert weights on
+CUDA and 22.13 GiB on CPU. This provisional eight-block configuration adds
+approximately 1.89 GiB of expert weights to CUDA while reducing the context
+pool by 98304 cells. The final placement will be selected after measuring this
+new headroom and discussing layer priorities.
 
 To list the buffer type names on the host, run a bogus override:
 
@@ -87,8 +89,8 @@ llama-server \
   --mmproj mmproj-Ornith-1.5-35B-A3B-f16.gguf \
   --chat-template-file chat-template.jinja \
   --spec-type draft-mtp \
-  --override-tensor "blk\\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32|33|34|35)\\.ffn_(up|down|gate|gate_up)_(ch|)exps=CPU" \
-  -ngl all -c 589824 -np 3 \
+  --override-tensor "blk\\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32)\\.ffn_(up|down|gate|gate_up)_(ch|)exps=CPU" \
+  -ngl all -c 491520 -np 3 \
   --cache-type-k f16 --cache-type-v f16 --no-mmap \
   --temp 1.0 --top-p 0.95 --top-k 20
 ```
@@ -97,10 +99,11 @@ llama-server \
 
 1. The 27.2 GiB Q6_K file and 0.84 GiB vision projector fit the 62 GiB RAM
    box with the selected expert offload.
-2. The context pool contains 589824 cells. Three slots receive 196608 cells
-   each, which is 64K below the model context per stream.
-3. `--n-cpu-moe` remains supported. This service prefers the explicit tensor
-   override because it preserves the final five expert blocks on CUDA.
+2. The provisional context pool contains 491520 cells. Three slots receive
+   163840 cells each, which is 96K below the model context per stream.
+3. `--n-cpu-moe` remains supported. This service currently prefers the
+   explicit tensor override because it preserves the final eight expert
+   blocks on CUDA pending the next layer-priority decision.
 4. The service uses the Qwen-Sharp template from
    `peculiar-ragdoll/Qwen-Sharp-Chat-Templates` through
    `scripts/chat-template.jinja`.