[Model][Engram] Share host tables across co-located DP replicas by default

[Juntian777]

Purpose

EngramConfig.dp_shared_memory (#56512) keeps one copy of the CPU-offloaded Engram tables per node instead of one per DP replica, with no per-step Engram DP collectives, but it had to be requested by hand and failed startup whenever its requirements were not met. This PR turns it on by default wherever it can work and falls back otherwise:

• dp_shared_memory defaults to unset and resolves to enabled when cpu_offload is on, data_parallel_size > 1 and elastic EP is off. An explicit true keeps the existing strict validation.
• Before the layers are built, can_share_engram_tables falls back to the existing DP head sharding (same host footprint, one table per node) when the node-local Engram DP group has size 1 (replicas not co-located) or /dev/shm is smaller than the full tables: 188.8 GiB for DeepSeek-V4.1-Flash, against Docker's default 64 MiB. Both cases log a warning_once pointing at --shm-size / --ipc=host. The check compares against the tmpfs total, which is identical on every rank of a node, so no collective is needed to keep the decision consistent.
• The load_format allowlist is dropped: dummy (fill_) and sharded_state (copy_) write identical bytes into the shared mapping on every rank, so the restriction guarded nothing.

Duplicate-work check (2026-09-19): #57195 shares host lookups for SP token owners, #56757 backs tables with disk files, #57491 keeps tables in host memory on ROCm. None changes the dp_shared_memory default or adds a /dev/shm fallback.

Test plan

.venv/bin/python -m pytest tests/distributed/test_engram_dp_shard.py \
  tests/test_config.py -q -k engram
.venv/bin/pre-commit run --files vllm/config/engram.py vllm/config/vllm.py \
  vllm/models/deepseek_v41/nvidia/engram.py vllm/models/deepseek_v41/nvidia/model.py \
  tests/test_config.py tests/distributed/test_engram_dp_shard.py
.venv/bin/pre-commit run mypy-3.12 --hook-stage manual --files vllm/config/engram.py \
  vllm/config/vllm.py vllm/models/deepseek_v41/nvidia/engram.py \
  vllm/models/deepseek_v41/nvidia/model.py
git diff --check

Test result

• 68 passed on 4xGB200, including the four-process shared-memory storage tests, the new default-resolution cases (offloaded DP on; no offload, DP=1 and elastic EP off) and the new /dev/shm-too-small fallback test.
• Pre-commit, mypy-3.12 and git diff --check passed.

Model evaluation

DeepSeek-V4.1-Flash (dba1be0a), one GB200 node, DEP4 (TP1 x DP4, EP), DSpark 5 tokens with adaptive verification, no --engram-config flag. GSM8K 5-shot via MQA, completions API, temperature=0, max_tokens=1024, n=1319.

Metric	Value
exact_match/flexible-extract	0.931
exact_match/strict-match	0.931
num_request_errors	0

This matches the 0.9303 recorded for the same model on main (2026-09-13). The engine log confirms the default took the shared path on its own: Resolved Engram configuration: EngramConfig(cpu_offload=True, embedding_across_dp=False, dp_shared_memory=True), Engram node-local DP group size: 4 (TP=1), and both tables reported 94.42 GiB shared across DP replicas with no fallback warning. On the node, /dev/shm held 190 GiB and free reported 191 GiB shared, one copy of the tables rather than four private pinned copies.

AI assistance was used for implementation, testing, and this draft (Claude Code). Human review of every changed line and independent execution of relevant tests are required before submission.

🤖 Generated with Claude Code

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @Juntian777.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[ZJY0516]

/ci run

[github-actions]

✅ @Juntian777, CI is now available for this PR.

• /ci run starts upstream CI; /amd-ci run starts AMD CI only.
• Your branch must contain every commit currently on its upstream target branch. Merge or rebase onto the latest target branch, then rerun the command. Append --allow-stale to a run command to test an outdated branch at your own risk.
• /ci retry retries failed jobs in the CI build for the current PR head. If the current head has no CI build, it starts a new CI build for the current head containing only jobs that failed in the latest earlier CI build for this PR.
• /amd-ci retry retries failed jobs in AMD CI for the current PR head. Use /amd-ci run when the current head has no AMD CI build.
• /ci cancel cancels scheduled or running CI builds for this PR branch; /amd-ci cancel does the same for AMD CI only.

[github-actions]

✅ Triggered Buildkite CI #90149 for commit d78f21af6bdf.

[Juntian777]

/ci run

[github-actions]

✅ Triggered Buildkite CI #90161 for commit e2b0f6b93bc6.