[hnyls2002]

• engram_gather takes sort_ids: it sorts the ids, reads the rows in address order and stores each row at its original position. Output is bitwise identical.
• EngramEmbedding turns it on for host tables only when CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES is 0, i.e. the GPU maps pinned host memory in its own page tables (PCIe hosts). On NVLink-C2C hosts the attribute is 1, reads go through the host page tables, sorting only adds work, and the path is unchanged. Device tables and HIP are unchanged. No new flag.
• Why: random rows spread over a host table far larger than the GPU's address-translation reach miss on nearly every row; the same rows read in address order share pages, which makes the host-table lookup several times faster on PCIe hosts.

No behavior change except the order of host-table reads on GPUs without host page tables.

Verification: bitwise-equal gather output; DeepSeek-V4.1-Flash DEP4 with host Engram tables on a PCIe host is faster in prefill, and on an NVLink-C2C host the path is not taken (throughput unchanged); GSM8K unchanged.

---

CI States

Latest PR Test (Base): 🚫 Run #39841033519
Latest PR Test (Extra): ❌ Run #39841032986
Latest PR Test (AMD ROCm 10): 🚫 Run #39841033377

[hnyls2002]

/rerun-test test_engram_history.py test_deepseek_v41_flash_gb300_dp2ep2.py test_deepseek_v41_flash_gb300_tp2ep2.py

[hnyls2002]

/tag-and-rerun-ci

[github-actions]

🚀 ubuntu-latest (1 test): ✅ View workflow run

cd test/ && python3 registered/unit/layers/test_engram_history.py

🚀 2-gpu-gb300 (2 tests): ✅ View workflow run

cd test/ && python3 registered/e2e/models/test_deepseek_v41_flash_gb300_dp2ep2.py
cd test/ && python3 registered/e2e/models/test_deepseek_v41_flash_gb300_tp2ep2.py