Conversation
timethink
pushed a commit
to timethink/sglang
that referenced
this pull request
Mar 9, 2025
hhhh1252023
pushed a commit
to hhhh1252023/sglang_public
that referenced
this pull request
Jun 27, 2026
Automated code for submitting mimo_v2_flash model requests
This was referenced Jul 31, 2026
efschu
pushed a commit
to efschu/htsglang
that referenced
this pull request
Aug 26, 2026
…f from a list THE QUESTION WAS ASKED OF MY OWN EARLIER WORK, AND THE HONEST ANSWER WAS NO. Three duck-typed attribute misses on the Unified family had been found one at a time -- sgl-project#832 `evictable_size` (killed all three ranks on W29), sgl-project#872 `_drain_storage_control_queues_local`, #872b `value`/`*lock_ref` on UnifiedTreeNode. All three were INSTANCE fixes. The two conformance suites from sgl-project#872/#872b were the first class-level checks in this tree, and measuring them shows they were not enough: * the sgl-project#872 suite reads probed names out of the SOURCE -- genuinely derived, but for ONE module (`hicache_flip_writeback`); * the #872b suite carries `expected_unreadable = {"UnifiedTreeNode"}` and pins `value` plus the `*lock_ref` trio. That is a LIST. Measured: `TreeNode` exposes SIX names `UnifiedTreeNode` lacks -- value, lock_ref, host_value, host_ref_counter, protect_host, release_host. The list covered TWO of six: exactly the ones that had already hurt. So the #872b half was a third instance fix in class clothing, and the fourth instance would have landed. A list extended at each new finding is a COUNTER. WHAT THIS IS INSTEAD. The surface is computed from the BASE at test time: whatever the non-Unified counterpart provides and the Unified one does not. Consumer modules are then scanned for duck-typed reads of those names. Nothing about the three known pains is privileged, and a new divergence enters the surface without anyone editing this file. Measured on `UnifiedRadixCache`: 17 names `RadixCache` provides that it lacks, 40 against `MambaRadixCache`. THE DISCRIMINATOR, WITHOUT WHICH IT CRIES WOLF. Not every `hasattr` on a diverging name is a cross-class probe. `session_radix_cache` guards `_session_leaves` with `hasattr` and then CREATES it (`_ensure_session_radix_state`) -- lazy init, entirely correct, and a false positive here. The rule that separates them, and it is derived rather than listed: a probe is a cross-class miss only if the probing module NEVER ASSIGNS the attribute itself. Without it this gate fires on working code, and a gate that fires on working code gets muted -- which is how the original condition survived three findings. WHAT IT FOUND THAT NO INSTANCE FIX EVER TOUCHED. Four sites are derived today. Three were known. The fourth is `UnifiedTreeNode.host_value`, duck-typed in `hicache_demotion.is_persistable` -- named as a landmine during #872b but covered by NO check until now. It remains UNREACHABLE (hicache_demotion is imported only by hiradix_cache.py, zero references from unified_radix_cache), and it is RECORDED rather than fixed, so it cannot become live unnoticed. `sgl-project#832`'s `evictable_size` correctly produces NO finding -- its consumer now reads `component_evictable_size_`, so the check confirms that fix still holds. THE NAMING TRAP, RECORDED because it is why two readers reached the same wrong conclusion independently: `Unified*` spans TWO unrelated subsystems. `UnifiedRadixCache`/`UnifiedTreeNode` are the radix TREE and are LIVE on this rig. The unified memory POOLS are gated behind `--enable-unified-memory`, which is False by default (server_args.py:1378), is set nowhere in the tree, and whose handler asserts `speculative_algorithm is None` while this rig always runs MTP. A shared prefix suggested a shared gate that does not exist, so this suite pairs only the TREE classes and says why. TESTS (hermetic; no boot; no card touched). CUDA_VISIBLE_DEVICES="" verified at the process, `nvidia-smi --query-compute-apps` empty throughout, cards 0/0/0. * CAN-FAIL ON A NAME THAT NEVER HURT, which is the point of the exercise: - un-recording `host_value` (no instance fix ever touched it) -> gate red; - injecting a BRAND-NEW divergent probe into real consumer code (`getattr(node, "protect_host", ...)` in hicache_demotion -- a name in no list anywhere) -> gate red AND names it in the failure message; - dropping the lazy-init discriminator -> 2 red, incl. the `_session_leaves` false positive returning. Green restored after each. If only the three known names had reddened, the answer to "does this cover the family" would have been no. * partitioned tier-2 gate: wide 3775 / narrow 291 / serial 353 passed, 24 exclusions reported, names agrees with the summary on every lane. TWO failures, neither mine -- this commit adds a single test file under mem_cache/ and touches no product code: - test_pp_admission_wraparound_never_blocks: PRE-EXISTING, proved earlier on the clean base 726f11e where that module fails TWO tests, i.e. MORE than here; - test_pp_admission_chain_flush_deadlock_795: PASSES in isolation on this tree (both modules together: 1 failed, 6 passed), so it is the order/contention shape the narrow lane's bounded worker count exists to avoid, already seen twice this session. * test/registered/unit/mem_cache: 2 failed / 1814 passed -- the known test_acceptance_emitters_758 RefillTiming pair; +6 over 1808 is this suite. WHERE THIS STRUCTURALLY DOES NOT REACH, named in the module rather than left to be discovered: factory indirection (a class produced by a builder, never named here); `getattr(obj, some_variable)` where the name is computed; registry or dispatch-table lookup, where neither side appears in the source; and a consumer that reads `obj.attr` directly and lets AttributeError fly -- out of scope by intent, because that failure is LOUD and therefore a different, better problem. The same methodological limit as the post-set enumeration and the sgl-project#880 reachability sweep: stated, not papered over. NOT ESTABLISHED. Whether the pairing itself is complete beyond the radix-tree module -- `test_every_unified_class_is_paired` enforces it for `unified_radix_cache`, but a Unified* class living in another module is outside that scan by construction.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.