Bump base to nightly-124154a8; vendor vllm#50693 to lift the #50298 pin

[fank]

What

Bump the base from nightly-6f91edf9 (07-29) to nightly-124154a8 (08-01) and vendor vllm#50693, which lifts the reason we were pinned.

Why the pin existed, and why it can go

PR #38 pinned us one nightly below vllm#50298, because that PR's warmup branch in models/deepseek_v4/nvidia/flashmla.py asserts self.topk_indices_buffer is not None. A DSpark drafter has no indexer buffer, so profile_run died with a bare AssertionError after the full ~32 min weight load — reported as vllm#50615.

vllm#50693 (opened 08-01, fixes #50615) makes the assert conditional:

-  assert self.topk_indices_buffer is not None
-  top_k = 0 if swa_only else self.topk_indices_buffer.shape[-1]
+  if swa_only:
+      top_k = 0
+  else:
+      assert self.topk_indices_buffer is not None
+      top_k = self.topk_indices_buffer.shape[-1]

Verified that this actually covers our checkpoint, which was the open question in #50615's own text — the fix only helps if the drafter is SWA-only. DeepSeek-V4-Flash-0731's compress_ratios has 46 entries for 43 hidden layers, and the trailing three (the MTP/DSpark layers) are 0. Since swa_only = compress_ratio <= 1, our draft layers take the fixed branch. A checkpoint whose draft layers compress would still trip it; that caveat is now in the Dockerfile comment.

What the bump buys

PR
#50298	~1.88x kernel — the thing we deliberately pinned away from
#49236	3.9% E2E TTFT (workspace reuse for eager break)
#48047	sparse-MLA q-head padding removed for FlashInfer >=0.6.14
#50312	448 MiB — still a no-op for us (spec-decode deployments need the buffer)

Also resets the nightly prune clock: only ~10 nightlies are retained (currently back to 07-23), so 6f91edf9 expires around mid-August.

Verification

• 124154a8 is a strict superset of the current pin — git merge-base --is-ancestor 6f91edf96 124154a8 passes (166 commits ahead), so this is not a release-branch-cut trap
• all five patches git apply --check -p1 clean against 124154a8: the four existing ones plus #50693
• the base still needs every existing patch — #46257, #48023 and #50686 are all still open upstream

Tripwire rewritten, not deleted

The old tripwire asserted that the warmup branch contains no assert on topk_indices_buffer. That is now the wrong invariant — post-#50693 the assert is expected, it just has to sit behind the swa_only guard. The new check asserts ordering:

guard = warmup.find("if swa_only:")
assert guard != -1, "swa_only guard missing -> vllm#50693 not applied"
first_assert = warmup.find("assert self.topk_indices_buffer is not None")
assert first_assert == -1 or first_assert > guard

Confirmed it discriminates — run against 124154a8 before and after applying the patch:

BEFORE patch:  FAIL: swa_only guard missing -> 50693 not applied
AFTER patch:   PASS: DSpark-safe

A tripwire that passes on both states would be decoration; this one was tested against both.

Not addressed here

• #50686 is still open, so the encoder patch keeps travelling with us.
• Tool-call grammar is unchanged. _any_tool_strict still gates the structural tag behind at least one strict: true tool, and pi sends strict: false, so DSML generation stays unconstrained.
• #48931 unfixed — START-token omission still leaks orphan invokes as content (the ?-stalls).
• reasoning_effort: "high" is still a no-op on this SHA; only "max" injects anything.

Cost

Five vendored patches, all against open PRs, on a base that expires in ~2 weeks. Each bump means re-verifying five patches and four tripwires. The real exit is the next stable release, which would carry #48993/#48317/#48957 and be a tag — this buys performance and time, not an escape from the cycle.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.