rocm: rebased on top of the current main, removed unnecessary (previously introduced) changes to ds4_cuda.cu

CONTRIBUTING.md

@@ -43,8 +43,10 @@ What they cover:
 - `--logprob-vectors`: compares local token bytes and top-logprob slices against
   official DeepSeek V4 Flash continuation vectors. This catches tokenizer,
   template, attention, and logits regressions.
-- `--long-context`: runs a long-context continuation regression from
-  `tests/long_context_security_prompt.txt`.
+- `--long-context`: runs a long-context story fact-recall regression from
+  `tests/long_context_story_prompt.txt`. The model must retrieve spelled-out
+  person-number assignments from a long prose prompt and return `Name=number`
+  lines that the test parses.
 - `--tool-call-quality`: exercises actual model behavior for DSML tool-call
   emission in both fast and exact paths.
 - `--metal-kernels`: isolated Metal kernel numeric checks.

README.md

@@ -18,6 +18,8 @@ This project would not exist without **llama.cpp and GGML**, make sure to read
 the acknowledgements section, a big thank you to Georgi Gerganov and all the
 other contributors.
 
+## Motivations
+
 Now, back at this project. Why we believe DeepSeek v4 Flash to be a pretty special
 model deserving a stand alone engine? Because after comparing it with powerful smaller
 dense models, we can report that: