fix(composes): drop GENESIS_ENABLE_P65 + P85 — out of sync with v7.69 dispatcher v2

Reported by @efschu in discussion #25 — Genesis v7.69 dispatcher v2
fires validator errors on our shipped env-var bundles:

  validator ERROR: P85 — missing required dependency: P85 requires 'P84'
                   to also be APPLY (currently SKIP)
  validator ERROR: P67 — conflict: P65 and P67 are both APPLY but
                   declared mutually exclusive — pick one
  validator ERROR: P67b — conflict: P65 and P67b are both APPLY but
                   declared mutually exclusive — pick one

The conflicts are real. Our shipped env vars in long-text.yml,
long-vision.yml, bounded-thinking.yml, and dual-turbo.yml all enabled
GENESIS_ENABLE_P65 + P67 + P85 simultaneously. v7.69 dispatcher v2
declares:

- **P65 vs P67/P67b**: P65 was the cudagraph-downgrade *workaround* for
  the spec-decode cascade bug (#40880). P67/P67b is the *proper Triton
  kernel fix* designed to replace it. Now that P67 is mature in v7.69,
  the dispatcher won't allow the workaround AND the fix simultaneously.
  Resolution: drop P65, keep P67. P67 supersedes P65 functionally.

- **P85 → P84**: P85 (hybrid fine-shadow prefix cache) requires P84
  (hash_block_size override, the actual root-cause fix for vllm#38182).
  We had P85 on but P84 unconfigured. Resolution: drop P85. P85 is
  opt-in optimization not load-bearing for correctness; enabling P84
  separately would need bench validation and is a future decision.

Verified on local test rig:
  Before: Genesis Results: 63 applied, 36 skipped, 0 failed
          (with 3 validator ERROR messages mid-log, non-blocking on
          our rig but confirmed blocking on @efschu's 2× 3080)
  After:  Genesis Results: 61 applied, 38 skipped, 0 failed
          0 validator errors, P67 active, P65 properly skipped,
          Application startup complete clean

Note: validator errors were non-blocking on our 3090 rig but @efschu
reports they break boot on 2× 3080 (different regime detection in
Genesis dispatcher → different error handling). Either way, they
indicate a config bug on our side and the fix is the same.

Comments added inline at each removal site so future maintainers
understand why the env vars were dropped (and that re-adding requires
checking dispatcher v2 conflict matrix).

Awaiting @efschu's confirmation log to see if this also unblocks his
2× 3080 boot, but the env-var conflict fix should land regardless.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
This commit is contained in:
noonghunna
2026-05-03 10:52:57 +00:00
co-authored by Claude Opus 4.7
parent 1bb85fadb8
commit a26e30b279
4 changed files with 14 additions and 8 deletions
@@ -133,7 +133,7 @@ services:
- GENESIS_ENABLE_P61B_STREAMING_OVERLAP=1
- GENESIS_ENABLE_P62_STRUCT_OUT_SPEC_TIMING=1
- GENESIS_ENABLE_P64_QWEN3CODER_MTP_STREAMING=1
- GENESIS_ENABLE_P65_TURBOQUANT_SPEC_CG_DOWNGRADE=1
# P65 dropped 2026-05-03 (mutually exclusive with P67/P67b in v7.69)
- GENESIS_ENABLE_P66_CUDAGRAPH_SIZE_FILTER=1
- GENESIS_ENABLE_P67_TQ_MULTI_QUERY_KERNEL=1
- GENESIS_ENABLE_P68_AUTO_FORCE_TOOL=1
@@ -143,7 +143,7 @@ services:
- GENESIS_PROFILE_RUN_CAP_M=4128
- GENESIS_ENABLE_P74_CHUNK_CLAMP=1
- GENESIS_ENABLE_P83=1
- GENESIS_ENABLE_P85=1
# P85 dropped 2026-05-03 (requires P84 which we don't enable in v7.69)
- GENESIS_ENABLE_P87=1
- GENESIS_ENABLE_P91=1
- GENESIS_ENABLE_P94=1
@@ -93,7 +93,7 @@ services:
- GENESIS_ENABLE_P61B_STREAMING_OVERLAP=1
- GENESIS_ENABLE_P62_STRUCT_OUT_SPEC_TIMING=1
- GENESIS_ENABLE_P64_QWEN3CODER_MTP_STREAMING=1
- GENESIS_ENABLE_P65_TURBOQUANT_SPEC_CG_DOWNGRADE=1
# P65 dropped 2026-05-03 (mutually exclusive with P67/P67b in v7.69)
- GENESIS_ENABLE_P66_CUDAGRAPH_SIZE_FILTER=1
- GENESIS_ENABLE_P67_TQ_MULTI_QUERY_KERNEL=1
- GENESIS_ENABLE_P68_AUTO_FORCE_TOOL=1
@@ -103,7 +103,7 @@ services:
- GENESIS_PROFILE_RUN_CAP_M=4128
- GENESIS_ENABLE_P74_CHUNK_CLAMP=1
- GENESIS_ENABLE_P83=1
- GENESIS_ENABLE_P85=1
# P85 dropped 2026-05-03 (requires P84 which we don't enable in v7.69)
- GENESIS_ENABLE_P87=1
- GENESIS_ENABLE_P91=1
- GENESIS_ENABLE_P94=1
@@ -154,7 +154,9 @@ services:
- GENESIS_ENABLE_P61B_STREAMING_OVERLAP=1
- GENESIS_ENABLE_P62_STRUCT_OUT_SPEC_TIMING=1
- GENESIS_ENABLE_P64_QWEN3CODER_MTP_STREAMING=1
- GENESIS_ENABLE_P65_TURBOQUANT_SPEC_CG_DOWNGRADE=1
# P65 (cudagraph downgrade workaround) dropped 2026-05-03 — Genesis
# v7.69 dispatcher v2 declares P65 mutually exclusive with P67/P67b
# (the proper Triton kernel fix that supersedes the workaround).
- GENESIS_ENABLE_P66_CUDAGRAPH_SIZE_FILTER=1
- GENESIS_ENABLE_P67_TQ_MULTI_QUERY_KERNEL=1
# P68/P69 — v7.65 raised default threshold 8000 → 50000 chars (~12.5K
@@ -167,7 +169,11 @@ services:
- GENESIS_PROFILE_RUN_CAP_M=4128
- GENESIS_ENABLE_P74_CHUNK_CLAMP=1
- GENESIS_ENABLE_P83=1
- GENESIS_ENABLE_P85=1
# P85 (hybrid fine-shadow prefix cache) dropped 2026-05-03 — Genesis
# v7.69 dispatcher v2 declares P85 requires P84 (hash_block_size
# override) which we don't enable. P85 is opt-in optimization not
# load-bearing for correctness; enabling P84 separately is a future
# decision (would need bench validation).
- GENESIS_ENABLE_P87=1
- GENESIS_ENABLE_P91=1
- GENESIS_ENABLE_P94=1
@@ -119,7 +119,7 @@ services:
- GENESIS_ENABLE_P61B_STREAMING_OVERLAP=1
- GENESIS_ENABLE_P62_STRUCT_OUT_SPEC_TIMING=1
- GENESIS_ENABLE_P64_QWEN3CODER_MTP_STREAMING=1
- GENESIS_ENABLE_P65_TURBOQUANT_SPEC_CG_DOWNGRADE=1
# P65 dropped 2026-05-03 (mutually exclusive with P67/P67b in v7.69)
- GENESIS_ENABLE_P66_CUDAGRAPH_SIZE_FILTER=1
- GENESIS_ENABLE_P67_TQ_MULTI_QUERY_KERNEL=1
- GENESIS_ENABLE_P68_AUTO_FORCE_TOOL=1
@@ -129,7 +129,7 @@ services:
- GENESIS_PROFILE_RUN_CAP_M=4128
- GENESIS_ENABLE_P74_CHUNK_CLAMP=1
- GENESIS_ENABLE_P83=1
- GENESIS_ENABLE_P85=1
# P85 dropped 2026-05-03 (requires P84 which we don't enable in v7.69)
- GENESIS_ENABLE_P87=1
- GENESIS_ENABLE_P91=1
- GENESIS_ENABLE_P94=1