Merge upstream→gfx11

[eble-amd]

Upstream and local Gemma 4 code had some similar changes to fix issues. Where there were conflicts, I kept the upstream version.

I'll record my manual testing in comments.

[eble-amd]

The Gemma 4 merge conflicts that I resolved involved #1016, so I ran the following local tests with and without --dynamic-lm-head-quantization int8:

With:

Arguments: --model cyankiwi/gemma-4-E4B-it-AWQ-INT4 --num-prompts 10 --max-model-len 4096 --kill-existing-vllm-processes --input-len 100 --output-len 200 --dynamic-lm-head-quantization int8 --target-gpu-memory-gb 20 --max-num-seqs 1 --synthetic-mm --synthetic-mm-width 1024 --synthetic-mm-height 800 --synthetic-mm-num-images 1 --synthetic-mm-base-items-per-request 1 --mm-encoder-attn-backend TRITON_ATTN -e TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 -e FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE -e TORCH_BLAS_PREFER_HIPBLASLT=1
Prefill 1423.23 tokens/s; TTFT 255 ms
Decode 57.2 tokens/s (TPOT 17.49 ms)
End-to-end latency 3735 ms (median)
Average CPU utilization 6%

Without:

Arguments: --model cyankiwi/gemma-4-E4B-it-AWQ-INT4 --num-prompts 10 --max-model-len 4096 --kill-existing-vllm-processes --input-len 100 --output-len 200 --target-gpu-memory-gb 16 --max-num-seqs 1 --synthetic-mm --synthetic-mm-width 1024 --synthetic-mm-height 800 --synthetic-mm-num-images 1 --synthetic-mm-base-items-per-request 1 --mm-encoder-attn-backend TRITON_ATTN -e TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 -e FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE -e TORCH_BLAS_PREFER_HIPBLASLT=1
Prefill 1403.39 tokens/s; TTFT 259 ms
Decode 49.5 tokens/s (TPOT 20.22 ms)
End-to-end latency 4282 ms (median)
Average CPU utilization 6%