feat(beam): Beam 1.0 model support for the 1.6.0 Beam release by luckyq · Pull Request #16702 · ai-dynamo/dynamo

luckyq · 2026-10-16T23:06:59Z

Summary

Brings the Dynamo Beam 1.0 model support to release/1.6.0-beam-1.0-501b-dev.1, as built into the 1.6.0 Beam release images. Counterpart of frontend-crates #430.

Three commits on top of main c5773baaa5 (the commit the release images were built from):

fix(vllm): cherry-pick of upstream fix(vllm): import engine-launch helpers lazily so workers start on new vLLM #16378 (lazy vLLM engine-launch imports), needed with the Beam vLLM engine. The upstream squash commit carries no DCO trailer, so this cherry-pick is re-authored with the original author as co-author.
feat(beam): the Beam source as one commit.
docs(rl): cherry-pick of upstream docs(rl): fix dead Miles link failing lychee #16573 (dead link that fails the lychee check on this base).
- Beam parser family on the unified v2 route; the parser starts in the channel an assistant prefill leaves open; only the model's stop-token text is kept on the Beam route.
- Beam turn grammar (structural tag) on every chat turn; forced tool choices stay on the Harmony wire; builtin_tools declared by name for the 1.0 template; response_format held in the final channel with held tool arguments framed.
- Beam route policy: non-compliant strict tools served as non-strict with one warning per request; preserve_thinking=false accepted as a no-op; allow_missing_thinking defaults to true.
- /v1/completions echoes the prompt with its logprobs.
- vLLM: chat_utils imported from its new location when needed.
- Cargo: dynamo-renderer 8.1.0 and dynamo-parsers-v2 0.10.0 pinned from frontend-crates at the Beam release revision (ef8f4013, the Request Documentation/Examples for Running DeepSeek R1 671B with Prefill-Decode Separation in AI-Dynamo #430 squash merge on that branch).

Base branch

The release branch was fast-forwarded to c5773baaa5 on 2026-10-17, the commit the release images were built from; this PR is the two commits above on top of it.

Validation

Source-identical to the mirror branch the release images were built from; only the crate pins and two comment lines differ.
On GB300 with the Beam 1.0 weights: partner verifier 109/0/1 on both vLLM and SGLang, BFCL 91.x vs the 87.5 reference, AIME 2026 vLLM 0.964 / SGLang 0.968, Beam recipes aggregated and disaggregated.
CI on this branch is the check for the public build against the public frontend-crates revision.

copy-pr-bot · 2026-10-16T23:07:04Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

dynamo-review-agent

Previously reported defects still present:

Original discussion: The echoed prompt logprobs still set text_offset to an empty vector while adding one entry per prompt token to the other parallel arrays. Both streaming and aggregated echo: true responses therefore return offsets only for generated tokens, so clients cannot associate echoed prompt tokens with positions. Populate cumulative prompt-token offsets before appending the shifted generated offsets.
Original discussion: This resolved defect remains at the reviewed head: echoed prompt logprobs still add entries to tokens, token_logprobs, and top_logprobs while setting text_offset to an empty vector, so echo: true responses expose non-parallel per-token arrays. Populate one cumulative prompt-text offset per echoed token before appending shifted generated offsets.

dynamo-review-agent · 2026-10-17T23:24:57Z

+
+
+def _resolve_engine_utils() -> tuple[Callable[..., Any], Callable[..., Any]]:
+    """Bind the vLLM engine-launch helpers this pool needs, failing with a clear message.


This private helper docstring duplicates the immediately preceding explanation and the function signature/body: _resolve_engine_utils already shows that it binds and returns the two helpers, raises a tailored error, and synchronizes the module attributes used by tests. It adds no lasting constraint beyond that nearby documentation.

🤖 AI Fix

Remove the redundant _resolve_engine_utils docstring.

dynamo-review-agent · 2026-10-17T23:24:57Z

+    #[test]
+    fn preserve_thinking_is_rejected_off_the_beam_route() {
+        let preprocessor = synthetic_preprocessor(false, HISTORY_TEMPLATE);
+        for value in [true, false] {


The true and false iterations cannot detect different regressions: the non-Beam production branch rejects solely on request.preserve_thinking.is_some(), without inspecting the Boolean value. Keep one supplied value to protect rejection of the field; the existing omitted case still protects acceptance when the field is absent.

🤖 AI Fix

Replace the [true, false] loop with one representative supplied preserve_thinking value.

dynamo-review-agent · 2026-10-17T23:24:57Z

+
+    #[test]
+    fn preserve_thinking_rejects_non_bool_values() {
+        for bad in [json!("true"), json!("yes"), json!(1), json!({})] {


"true" and "yes" are both JSON strings and exercise the identical Serde string-to-Option<bool> type rejection; their contents are never interpreted. Retaining one string plus the numeric and object cases preserves the distinct invalid JSON-type coverage.

🤖 AI Fix

Remove either json!("true") or json!("yes") from the invalid-value cases.

dynamo-review-agent

Previously reported defects still present:

Original discussion: Both JSON string cases remain in the invalid-value table even though Serde rejects them through the same string-to-Boolean type path; retain one string alongside the numeric and object cases.
Original discussion: The [true, false] loop remains, although the non-Beam branch rejects solely on preserve_thinking.is_some(); retain one supplied value and the omitted case.
Original discussion: The echoed prompt logprobs still add one entry per prompt token to tokens, token_logprobs, and top_logprobs while constructing text_offset as an empty vector. After CompletionEcho::prepend_to merges generated offsets, both streaming and aggregated echo: true responses expose non-parallel per-token arrays and omit positions for every echoed prompt token. Populate cumulative prompt-token offsets before appending the shifted generated offsets.
Original discussion: The echoed prompt still adds one entry per prompt token to tokens, token_logprobs, and top_logprobs while constructing text_offset as an empty vector, leaving the response’s parallel per-token arrays misaligned. Populate cumulative prompt-token offsets before appending shifted generated offsets.
Original discussion: The redundant _resolve_engine_utils docstring remains at components/src/dynamo/vllm/embedding_worker_processes.py:44-48, duplicating the preceding module comment and the helper's signature/body.
Original discussion: The same redundant explanation that either parser slot selects Beam remains, moved to the test comment at lib/llm/src/preprocessor/structural_tag.rs:1259-1260; the assertions already show both parser-slot cases.
Original discussion: Verified still present: echoed prompt logprobs populate tokens, token_logprobs, and top_logprobs per prompt token but leave text_offset empty. After folding, echo: true responses expose non-parallel arrays and offsets only for generated tokens. Populate cumulative prompt-token offsets before appending shifted generated offsets.

dynamo-review-agent · 2026-10-18T20:41:26Z

        }
    }

+    /// `retain_model_stop_token_text` keeps the model's own EOS text (what a parser like


This test doc comment repeats the function name and the immediately visible EOS/user-stop/string cases without preserving an additional constraint.

🤖 AI Fix

Remove the redundant test doc comment.

dynamo-review-agent · 2026-10-18T20:41:27Z

+        assert_eq!(tools[0].function.strict, None);
+        assert_eq!(tools[1].function.strict, Some(true));
+        assert!(ValidateRequest::validate(&beam).is_ok());
+        // Idempotent.


This comment merely labels the repeated call and zero-result assertion as idempotent; the code already expresses that property.

🤖 AI Fix

Remove the Idempotent comment.

dynamo-review-agent · 2026-10-18T20:41:27Z

+    }
+}
+
+/// Apply a non-default framing before the parser consumes its first byte.


This private-helper doc comment restates apply_argument_framing and its explicit non-default check without recording any non-obvious requirement.

🤖 AI Fix

Remove the redundant helper doc comment.

dynamo-review-agent · 2026-10-18T20:41:27Z

+                },
+                ArgumentFraming::SchemaConstrained,
+            ),
+            (


The frames_answer: true row cannot distinguish a framing regression from the retained frames_answer: false row: argument_framing matches BeamTurnGrammar { .. } and deliberately ignores this payload. The later assertions in this same test already cover the payload's distinct structured-response behavior, so keeping one Beam row here still protects the SchemaConstrained mapping.

🤖 AI Fix

Remove the BeamTurnGrammar { frames_answer: true } entry from the argument_framing table.

dynamo-review-agent · 2026-10-18T20:41:27Z

+            let echo = self.echo.as_mut().expect("checked by wants_echo");
+            let logprobs = match &prompt_logprobs {
+                Some(payload) if enable_logprobs => Some(echo.logprobs(payload, return_as_ids)),
+                None if enable_logprobs => {


Supported workers attach engine_data.prompt_logprobs to the final chunk (the vLLM handler explicitly captures the early engine payload and moves it there), but this path treats their absence on the first generated chunk as final and immediately records the choice in emitted. Consequently, ordinary multi-chunk echo: true requests with logprobs return only echoed text and permanently drop the requested prompt logprobs. Preserve streaming order while waiting for the prompt payload, or make the backend provide it on the first emitted choice chunk, and only then mark the echo emitted.

🤖 AI Fix

Buffer an echo + logprobs choice until its prompt-logprobs payload is available, or attach that payload to its first backend chunk, and mark the choice emitted only after constructing the complete echo.

…e engine reports them Problem (review on #16702): vLLM and SGLang attach the prompt logprobs to the final chunk. Echoing everything on the first chunk therefore dropped the prompt logprobs of every `echo: true` + `logprobs` request; echoing everything on the chunk that carries them put generated text before the prompt for streaming clients. Fix: the echo now travels in two parts. The prompt text rides the first chunk of each choice; the prompt logprobs ride the chunk that carries the engine's payload (text empty when it already went out). The unary aggregator merges the parts; `fold_echo` prepends whatever a chunk carries. If no payload arrives by the choice's end, the text-only echo stands and a warning is logged. Also from the review: one representative value in the preserve_thinking rejection tests, the duplicate `frames_answer: true` framing row dropped, and four comments that restated their code removed. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: doxu <doxu@nvidia.com>

dynamo-review-agent · 2026-10-18T21:10:44Z

Previously reported defects still present:

Original discussion: The private _resolve_engine_utils docstring remains at line 44 and still duplicates the preceding explanation plus the function signature and body; remove it as previously requested.
Original discussion: The reported echo: true logprob defect remains: EchoState::logprobs still creates prompt entries in tokens, token_logprobs, and top_logprobs while leaving text_offset empty, so folded responses expose non-parallel per-token arrays. Populate cumulative prompt-token offsets before appending shifted generated offsets.
Original discussion: The revised path no longer drops delayed prompt logprobs, but it still violates streaming order: generated-token logprobs are emitted in earlier SSE chunks, while the prompt logprobs are prepended only within the final chunk that carries the engine payload. Concatenating streamed logprob entries therefore yields completion-prefix, prompt, completion-tail instead of prompt followed by completion. Buffer echo + logprobs choice chunks until the prompt payload arrives, or arrange for that payload to be attached before the first generated chunk.
Original discussion: The revised separate text_emitted/logprobs_emitted tracking no longer drops a final prompt-logprobs payload, but it still does not preserve the requested streaming association: supported workers deliver that payload on the final chunk, where fold_echo prepends prompt-token logprobs to that chunk's generated-token logprobs while its text contains only the final generated delta. Ordered SSE logprob events therefore describe generated-prefix tokens before prompt tokens even though visible text is prompt-first. Buffer until the payload is available or attach it to the first emitted choice chunk as originally requested.
Original discussion: Verified still present: EchoState::logprobs adds one entry per prompt token to tokens, token_logprobs, and top_logprobs but initializes text_offset as empty. After folding, echo: true responses expose non-parallel per-token arrays and omit offsets for every echoed prompt token. Populate cumulative prompt-token offsets before appending shifted generated offsets.

lib/llm/src/preprocessor.rs:3529 — The pinned renderer explicitly allows a legacy builtin_tools declaration object whose name is absent from tools and appends that declaration to this render-only view. The rest of Beam preprocessing still builds beam_turn_structural_tag and parser tool definitions from the original request, so the prompt advertises a callable tool that the default turn grammar does not offer (and the parser lacks its schema). Such legacy built-ins therefore cannot be generated successfully despite being accepted. Use one effective tool set for rendering, grammar construction, and parsing.

🤖 AI Fix

Resolve Beam built-in declarations into a local effective-tools request view and reuse that view for template rendering, turn-grammar construction, and parser tool definitions.

…w vLLM (#16378) (cherry picked from commit ac00e43) Co-authored-by: Pei Li <peili@nvidia.com> Signed-off-by: doxu <doxu@nvidia.com> Co-Authored-By:

Problem: the 1.6.0 Beam release images were built from the Dynamo Beam source on a private branch, which the public release branch does not carry. Root cause: Beam support was developed on a private branch while the model was unreleased; the frontend crates it needs (dynamo-renderer 8.1.0, dynamo-parsers-v2 0.10.0) are on the frontend-crates `release/1.6.0-beam-1.0-501b-dev.1` line, not on crates.io. Fix: bring the Beam source onto the release branch as one commit. - Beam parser family served through the unified v2 route; the parser starts in the channel an assistant prefill leaves open; only the model's stop-token text is kept on the Beam route. - Beam turn grammar (structural tag) on every chat turn, with forced tool choices kept on the Harmony wire; `builtin_tools` declared by name for the 1.0 template; response_format held in the final channel and held tool arguments framed. - Beam route policy: non-compliant `strict` tools served as non-strict with one warning per request; `preserve_thinking=false` accepted as a no-op; `allow_missing_thinking` defaults to true. - /v1/completions echoes the prompt with its logprobs. - vLLM: import chat_utils from its new location when needed. - Cargo: pin dynamo-renderer 8.1.0 and dynamo-parsers-v2 0.10.0 from the public frontend-crates repository at the Beam release revision. Signed-off-by: doxu <doxu@nvidia.com> Co-Authored-By:

Signed-off-by: Nate Mailhot <nmailhot@nvidia.com> (cherry picked from commit d806094) Co-Authored-By:

…replies Problem (review on #16702): with `echo: true`, batch completions remapped `choice.index` by `prompt_idx * n` but left `echo.choice_index` as the engine saw it, so the echo of every prompt after the first found no choice; and when the engine reported no prompt logprobs the echo rode the choice's final chunk, so a streaming client saw generated text before the prompt. Fix: - Remap `echo.choice_index` with the same offset as the choices. - Echo on the first chunk of each choice: with prompt logprobs when that chunk carries them (engines report them with the first token), text only otherwise. Test updated for the new order. - Drop three doc comments that only restated the code. Signed-off-by: doxu <doxu@nvidia.com> Co-Authored-By:

…heck Problem: `dynamo-truthy`'s repository scan (`no_bool_parse_forks`) fails CI on the hand-rolled `"0" | "false" | "off" | "no"` match that reads `DYN_BEAM_TURN_GRAMMAR`. Fix: use `dynamo_runtime::config::env_is_falsey` on the variable; unset keeps the grammar on, as before. Signed-off-by: doxu <doxu@nvidia.com> Co-Authored-By:

frontend-crates #430 squash-merged into `release/1.6.0-beam-1.0-501b-dev.1` as ef8f4013; move dynamo-renderer 8.1.0 and dynamo-parsers-v2 0.10.0 from the PR head to that commit and refresh the three lock files. Signed-off-by: doxu <doxu@nvidia.com> Co-Authored-By:

…e engine reports them Problem (review on #16702): vLLM and SGLang attach the prompt logprobs to the final chunk. Echoing everything on the first chunk therefore dropped the prompt logprobs of every `echo: true` + `logprobs` request; echoing everything on the chunk that carries them put generated text before the prompt for streaming clients. Fix: the echo now travels in two parts. The prompt text rides the first chunk of each choice; the prompt logprobs ride the chunk that carries the engine's payload (text empty when it already went out). The unary aggregator merges the parts; `fold_echo` prepends whatever a chunk carries. If no payload arrives by the choice's end, the text-only echo stands and a warning is logged. Also from the review: one representative value in the preserve_thinking rejection tests, the duplicate `frames_answer: true` framing row dropped, and four comments that restated their code removed. Signed-off-by: doxu <doxu@nvidia.com> Co-Authored-By:

dynamo-review-agent

Previously reported defects still present:

Original discussion: Beam rendering still accepts and advertises a legacy builtin_tools declaration absent from the request's tools, while grammar construction and parser definitions use only the original tools. Build one effective tool set and reuse it for rendering, turn-grammar construction, and parsing.
Original discussion: Streaming echo + logprobs still emits generated-token logprobs before the prompt logprobs that supported workers attach to the final chunk, producing completion-prefix, prompt, completion-tail order. Buffer these choice chunks until the prompt payload arrives, or attach the prompt payload before the first generated chunk.
Original discussion: The previously reported explanation that either parser slot selects Beam remains, moved to lib/llm/src/preprocessor/structural_tag.rs:1259-1260; the adjacent test cases already express both parser-slot paths.
Original discussion: The redundant _resolve_engine_utils docstring remains at components/src/dynamo/vllm/embedding_worker_processes.py:44-48, duplicating the preceding compatibility explanation and the helper's signature/body.
Original discussion: Verified still present: Beam's builtin_tools_view changes only the request view used for template rendering, while turn-grammar construction and parser tool definitions still use the original request tools. A legacy built-in accepted and advertised in the prompt can therefore be absent from the grammar and parser schema. Resolve one effective tool set and reuse it for rendering, grammar construction, and parsing.
Original discussion: Verified still present: EchoState::logprobs creates prompt entries in tokens, token_logprobs, and top_logprobs but leaves text_offset empty, so echo: true responses expose non-parallel per-token arrays and omit offsets for echoed prompt tokens. Populate cumulative prompt-token offsets before appending shifted generated offsets.
Original discussion: Verified still present: supported workers provide prompt logprobs on the final chunk after generated-token logprobs have already streamed, producing the client-visible order generated prefix, prompt, generated tail instead of prompt followed by completion. Buffer echo + logprobs choice chunks until the prompt payload arrives, or attach the prompt payload before the first generated chunk.
Original discussion: Verified still present: supported workers provide prompt logprobs on the final chunk, while earlier generated-token logprobs are already streamed. The final chunk prepends prompt logprobs only within that chunk, so SSE order is generated-prefix, prompt, generated-tail rather than prompt followed by completion. Buffer echo + logprobs chunks until the prompt payload arrives or attach that payload before the first generated chunk.
Original discussion: EchoState::logprobs still creates prompt entries in tokens, token_logprobs, and top_logprobs while leaving text_offset empty, so echoed completion responses expose non-parallel per-token arrays. Populate a cumulative offset for every prompt token before appending shifted generated offsets.

dynamo-review-agent · 2026-10-18T22:10:52Z

+        let tools_before = serde_json::to_value(&request.inner.tools).unwrap();
+        let rendered = preprocessor.apply_template(&request).unwrap().unwrap();
+        assert_eq!(rendered.as_str(), "SYS:bash:string;DEV:weather;");
+        // The request's real tools and kwargs are untouched.


This test comment only narrates the two immediately following assertions, which directly verify that tools and builtin_tools remain unchanged; it preserves no additional invariant.

🤖 AI Fix

Remove the redundant test comment.

…ompletions until their prompt logprobs Problem (review on #16702): - A legacy `builtin_tools` declaration object absent from `tools` was rendered into the prompt but unknown to the turn grammar and the parser, so the model could be shown a tool it could not call. - With `echo: true` and `logprobs`, generated logprobs streamed before the prompt logprobs that vLLM and SGLang attach to the final chunk. Fix: - `declare_legacy_builtin_tools_for_beam` adds such declarations to the request's `tools` on the Beam route, before validation, in both the HTTP handler and the in-process preprocessor entry; the renderer then sees them as declared and the grammar and parser build from the same list. - The completions delta generator holds a choice's generated text and logprobs while its prompt logprobs are pending and releases them, prompt positions first, on the chunk that carries the payload. - Two test comments that restated their assertions removed. Signed-off-by: doxu <doxu@nvidia.com>

dynamo-review-agent

Previously reported defects still present:

Original discussion: The previously reported echo-logprobs defect remains: EchoState::logprobs still creates one tokens, token_logprobs, and top_logprobs entry per prompt token while leaving text_offset empty, so echoed completion responses expose non-parallel per-token arrays. Populate a cumulative prompt-text offset for every echoed token before merging generated offsets.
Original discussion: The redundant _resolve_engine_utils docstring remains at lines 44–48, duplicating the preceding module comment and the helper’s signature and body.
Original discussion: The reported echo: true logprob defect remains: EchoState::logprobs populates tokens, token_logprobs, and top_logprobs for each prompt token while leaving text_offset empty, so echoed responses expose non-parallel per-token arrays. Populate cumulative prompt-token offsets before merging generated offsets.

dynamo-review-agent · 2026-10-18T22:38:04Z

+            Some(&plain),
+        );
+        assert!(!prefill.unified && !prefill.response);
+        // Another family's prompt never sets the Beam response prefill.


This test comment merely narrates the immediately following Qwen-family input and assert!(!prefill.response); it preserves no additional invariant.

🤖 AI Fix

Remove the redundant test comment.

dynamo-review-agent · 2026-10-18T22:38:05Z

+        assert_eq!(applied, GuidedToolConstraint::None);
+        assert!(installed_structural_tag(&common_request).is_none());
+
+        // Other parsers see no Beam grammar.


This comment only labels the hermes case whose subsequent assertion already shows that a non-Beam parser installs no Beam structural tag.

🤖 AI Fix

Remove the redundant test comment.

dynamo-review-agent · 2026-10-18T22:38:05Z

+                .iter_mut()
+                .find(|choice| choice.index == index)
+        {
+            if prompt_logprobs.is_none() && finish_reason.is_none() {


For streaming completions with both echo and logprobs, supported vLLM/SGLang workers attach prompt logprobs to the final chunk, so this branch removes every generated text/logprobs delta and accumulates it until generation finishes. The response therefore behaves effectively as unary after the initial echoed-prompt chunk, increasing time-to-first-generated-token to full generation time and retaining the complete output per choice in memory. Make prompt logprobs available on the first emitted choice chunk so subsequent generated deltas can stream normally.

🤖 AI Fix

Attach the engine's prompt-logprobs payload to each choice's first emitted chunk and emit subsequent generated text and logprobs without accumulating them in EchoState::held.

Signed-off-by: neilkg <33635204+neilkg@users.noreply.github.com>

Ports the Beam 1.0 support merged into release/1.6.0-beam-1.0-501b-dev.1 (#16702) onto main, pinned to the published crates instead of the release line: dynamo-parsers-v2 0.7.29 (the Beam unified family, frontend-crates #438), dynamo-renderer 8.1.0 (the Beam builtin_tools rendering view) and dynamo-tokenizers 3.1.2 (required by renderer 8.1.0). (cherry picked from commit b90142f) Signed-off-by: doxu <doxu@nvidia.com> Signed-off-by: Nate Mailhot <nmailhot@nvidia.com> Signed-off-by: neilkg <33635204+neilkg@users.noreply.github.com> Co-authored-by: Pei Li <peili@nvidia.com> Co-authored-by: Nate Mailhot <nmailhot@nvidia.com> Co-authored-by: neilkg <33635204+neilkg@users.noreply.github.com> Signed-off-by: doxu <doxu@nvidia.com>

luckyq requested review from a team as code owners October 16, 2026 23:07

github-actions Bot added feat documentation Improvements or additions to documentation backend::vllm Relates to the vllm backend deployment::k8s Relates to dynamo deployment in kubernetes backend::sglang Relates to the sglang backend backend::trtllm Relates to the trtllm backend labels Oct 16, 2026

dynamo-review-agent Bot reviewed Oct 17, 2026

View reviewed changes

dynamo-review-agent Bot reviewed Oct 18, 2026

View reviewed changes

luckyq and others added 7 commits October 18, 2026 15:03

fix(vllm): import engine-launch helpers lazily so workers start on ne…

5c88c8b

…w vLLM (#16378) (cherry picked from commit ac00e43) Co-authored-by: Pei Li <peili@nvidia.com> Signed-off-by: doxu <doxu@nvidia.com> Co-Authored-By:

docs(rl): fix dead Miles link failing lychee (#16573)

873912a

Signed-off-by: Nate Mailhot <nmailhot@nvidia.com> (cherry picked from commit d806094) Co-Authored-By:

luckyq force-pushed the luckyq/beam-1.0-release branch from 2eb9443 to 1b4e060 Compare October 18, 2026 22:03

dynamo-review-agent Bot reviewed Oct 18, 2026

View reviewed changes

ynpandey-nv approved these changes Oct 19, 2026

View reviewed changes

neilkg added 2 commits October 19, 2026 12:11

fix(beam): honor ignore_eos for unconstrained chat (#16784)

5ecbdb2

Signed-off-by: neilkg <33635204+neilkg@users.noreply.github.com>

fix(beam): reject unsupported reasoning-off requests (#16785)

d4b73de

Signed-off-by: neilkg <33635204+neilkg@users.noreply.github.com>

ynpandey-nv merged commit b90142f into release/1.6.0-beam-1.0-501b-dev.1 Oct 19, 2026
30 checks passed

ynpandey-nv deleted the luckyq/beam-1.0-release branch October 19, 2026 22:01

luckyq mentioned this pull request Oct 22, 2026

feat(beam): Beam 1.0 model support on main #17072

Merged



		def _resolve_engine_utils() -> tuple[Callable[..., Any], Callable[..., Any]]:
		"""Bind the vLLM engine-launch helpers this pool needs, failing with a clear message.

Conversation

luckyq commented Oct 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Base branch

Validation

Uh oh!

copy-pr-bot Bot commented Oct 16, 2026

Uh oh!

dynamo-review-agent Bot left a comment

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot Oct 17, 2026

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot Oct 17, 2026

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot Oct 17, 2026

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot left a comment

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot Oct 18, 2026

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot Oct 18, 2026

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot Oct 18, 2026

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot Oct 18, 2026

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot Oct 18, 2026

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot commented Oct 18, 2026

Uh oh!

dynamo-review-agent Bot left a comment

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot Oct 18, 2026

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot left a comment

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot Oct 18, 2026

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot Oct 18, 2026

Choose a reason for hiding this comment

Uh oh!

dynamo-review-agent Bot Oct 18, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

luckyq commented Oct 16, 2026 •

edited

Loading