Repository navigation
Codex/kluster integration - #1
Merged
Merged
Conversation
Make the QWEN (v1) loader rank-aware (rank-zero token embedding and output norm, per-owned-layer fused QKV with bias and gated FFN weights on owning ranks) and add build_qwen_segmented: rank-zero embedding graph, per-owned-layer NeoX-RoPE attention and SiLU parallel-FFN segments, and a rank-zero head graph, replacing the obsolete single-rank graph. Advertise qwen_segmented_v1. Co-authored-by: Cursor <cursoragent@cursor.com>
Make the PHI2 loader rank-aware (rank-zero token embedding and output norm/weight/bias on rank zero, per-owned-layer QKV with biases - fused or separate - output projection with bias and gated FFN weights with biases on owning ranks) and add build_phi2_segmented: rank-zero embedding graph, per-owned-layer partial-rotary attention with Q pre-scale, GELU FFN and parallel residual segments, and a rank-zero head graph with output bias, replacing the obsolete single-rank graph. Advertise phi2_segmented_v1. Co-authored-by: Cursor <cursoragent@cursor.com>
Make the PLAMO loader rank-aware (rank-zero token embedding and output norm/weight on rank zero, per-owned-layer Q/K/V and gated FFN weights on owning ranks) and add build_plamo_segmented: rank-zero embedding graph, per-owned-layer NORM-RoPE attention and SiLU parallel-FFN segments with parallel residual, and a rank-zero head graph, replacing the obsolete single-rank graph. Advertise plamo_segmented_v1. Co-authored-by: Cursor <cursoragent@cursor.com>
Make the CODESHELL loader rank-aware (rank-zero token embedding and output norms with biases, per-owned-layer QKV with bias, output projection with bias and GELU FFN weights with biases on owning ranks) and add build_codeshell_segmented: rank-zero embedding graph, per-owned-layer NeoX-RoPE attention and GELU FFN segments, and a rank-zero head graph, replacing the obsolete single-rank graph. Advertise codeshell_segmented_v1. Co-authored-by: Cursor <cursoragent@cursor.com>
Make the ORION loader rank-aware (rank-zero token embedding and output norms with biases on rank zero, per-owned-layer Q/K/V and gated FFN weights with norm biases on owning ranks) and add build_orion_segmented: rank-zero embedding graph, per-owned-layer NORM-RoPE attention and SiLU parallel-FFN segments, and a rank-zero head graph with biased output norm, replacing the obsolete single-rank graph. Advertise orion_segmented_v1. Co-authored-by: Cursor <cursoragent@cursor.com>
Make the INTERNLM2 loader rank-aware (rank-zero token embedding and output norm/weight on rank zero, per-owned-layer Q/K/V, output projection and gated FFN weights on owning ranks) and add build_internlm2_segmented: rank-zero embedding graph, per-owned-layer NORM-RoPE attention and SiLU parallel-FFN segments with RMS norms, and a rank-zero head graph, replacing the obsolete single-rank graph. Advertise internlm2_segmented_v1. Co-authored-by: Cursor <cursoragent@cursor.com>
Give MiniCPM (v1) a dedicated rank-aware loader separated from the shared MiniCPM3 MLA case (rank-zero token embedding, optional/tied output and output norm on rank zero, per-owned-layer Q/K/V with optional biases, output projection and gated FFN weights on owning ranks) and add build_minicpm_segmented: rank-zero embedding graph with 12x embedding scaling, per-owned-layer NORM-RoPE attention and SiLU parallel-FFN segments with depth-scaled residuals, and a rank-zero head graph, replacing the obsolete single-rank graph. Advertise minicpm_segmented_v1. Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
…atches - build_tool_parser_json_native: for tool_choice auto, add a content-only alternative so complete output that mentions a tool marker but is not a valid tool call degrades to plain content instead of failing the parse (server 500). tool_choice required keeps the strict tools-only path. - analyze_template: collect preserved tokens after the workaround patches run so tokenizer preservation agrees with patched tool/reasoning markers (e.g. DeepSeek-R1-Distill-Qwen). - Extend prima-chat-parser-contract-test with regression coverage for both behaviors (verified to fail without each fix).
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
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.