[jayinsine]

Problem

#167 declared a DeepSeek V4.1 top-p range of [0.95, 1]. --vocabulary-parallel-sampling supports only unfiltered target distributions (model-preparation.md: top-k, top-p, and min-p "must remain disabled"), and PreparedVocabularySoftmax marks a row invalid when top-p ≠ 1. Packages exported with the option after #167 accept top_p: 0.95, and every speculative round of that request then fails verification:

HTTP 500  prepared model reported device error 6   (INVALID_SPECULATION)

With --mtp-draft-tokens 0, the same package serves top_p: 0.95. Packages exported before #167 aren't affected because they clamp top-p to 1.

Change

• SamplingDefaults.unfiltered() returns the policy without top-k, top-p, and min-p ranges, and rejects defaults that enable those filters.
• DeepSeek V4.1's PreparationConfig.sampling_defaults packages that policy when vocabulary-parallel sampling is enabled, and __post_init__ checks it before any GPU work.
• As with any absent range, requests are clamped to the default (top-p 1).
• Other ranges, such as repetition_penalty [1, 1.1], are unchanged. Packages exported without the option keep the top-p range.
• No kernel, Helios, or package-contract changes. docs/model-preparation.md and docs/api.md describe the packaged policy.

Validation

Unit tests:

• New: the unfiltered policy keeps non-filter ranges (top-k −1 and 0) and rejects enabled filter defaults.
• New: the DeepSeek export adapter packages the expected ranges, with and without --vocabulary-parallel-sampling. This test fails without the bootstrap change.
• Affected suites (test_frontend_metadata, test_declarations, test_planning, test_speculative_graph_budget, test_deepseek41_frontend, test_native_recipe_contract, test_runtime_package; Heliopause built with qualification): 296 passed, 37 skipped, 11 failed.
• The same 11 fail on cc696a6b without this change:
  • five need the Kimi K3 checkpoint
  • two are the deepseek41 and muse_spark cases of test_default_mode_uses_manually_declared_model_card_sampling
  • two are runtime collective-ladder cases in test_every_recipe_passes_collective_policy_to_preparation
  • three are graph-budget cases

GPU: TP4 on four RTX PRO 6000 Blackwell GPUs.

• Export: a derivative of the qualified DeepSeek-V4.1-Flash-1017 revision, with --max-concurrency 32 --mtp-draft-tokens 7 --vocabulary-parallel-sampling --draft-head-quantization nvfp4 --markov-head-quantization fp8. Reuse retained warmup mappings across compiled warmup #154 was applied locally so the C32 export fits; it is not part of this PR. Served with flare run.
• flare inspect: the new package has no top-p range and keeps the repetition-penalty range. A package exported from 902c88e2 shows top-p [0.95, 1].
• top_p: 0.95: chat requests at C1 and C8 plus one streamed request all returned 200, with 3.4–3.8 accepted drafts per verification round. The 902c88e2 package returned device error 6 for the same request.
• Functional and speed: API checks passed 34/34, and decode throughput was unchanged against the previous export.

Limitations

• On these packages, a request for top-p below 1 is served at top-p 1, as before Add a hard reasoning-token budget and DeepSeek V4.1 sampling ranges #167, rather than with a filtered distribution. Supporting top-p in vocabulary-parallel verification would mean changing its sampling kernels; this PR doesn't attempt that.
• Only single-host TP4 was exercised.

Prepared with an AI agent (Claude Opus 5.5); see the Design-Signed-Off-By trailer. It doesn't touch CUDA, Triton, or GPU optimization code (design-review principle 14).

[original-el8]

Merged #208 (5e356e6) addresses the unsupported DeepSeek sampling range. It removes the top-p range from the DeepSeek V4.1-1017 declaration and rejects filtered defaults/ranges during vocabulary-parallel planning. This PR's expectation that non-vocabulary-parallel exports retain the 0.95–1 range also differs from #208's model-card correction.

Marked triage:superseded: close unless a distinct unsupported-filter case remains on main. Previously exported packages retain their stored sampling ranges and require re-export; merging another source patch would not update those packages.

[jayinsine]

Closing as superseded by #208. On main, DeepSeek V4.1-1017 declares no top-p range and vocabulary-parallel planning rejects filtered sampling defaults and ranges, which covers the device error 6 failure this PR addressed. Packages exported before #208, including ones built with this patch applied, keep their stored ranges until re-export.