Fix DFlash sglang target backend for sglang 0.5.13+ scheduler/batch API

[violet73]

Updated the dflash_target_model.py to replace deprecated methods and adjust input handling for batch processing.

Motivation

The DFlash sglang target backend (SGLangDFlashTargetModel in
specforge/modeling/target/dflash_target_model.py) still uses the pre-0.5.13
Req/ScheduleBatch scheduler API. On current sglang (0.5.13/0.5.14) an online
DFlash run crashes at the first generate_dflash_data call, before any training
step. This PR updates the request/batch construction to the new API, mirroring
the changes already made to the EAGLE3 target backend (eagle3_target_model.py).

Modifications

All in specforge/modeling/target/dflash_target_model.py
(SGLangDFlashTargetModel only; the hf backend is untouched):

• Imports: add from array import array; import the module-level
  prepare_mlp_sync_batch_raw from
  sglang.srt.managers.scheduler_components.dp_attn; drop the now-unused
  from sglang.srt.managers.scheduler import Scheduler.
• generate_dflash_data — Req construction: set
  full_untruncated_fill_ids / fill_len / extend_input_len /
  logprob_start_len instead of the removed fill_ids.
• _extend — batch prep: per req call init_next_round_input(tree_cache)
  then set_extend_range(len(prefix_indices), len(full_untruncated_fill_ids))
  (no prefix-cache reuse); capture input_lens before prepare_for_extend.
• _extend — mlp sync: call the module-level prepare_mlp_sync_batch_raw
  with attn_cp_size, without the removed spec_algorithm /
  speculative_num_draft_tokens.
• _extend — forward: materialize prefill_input_ids_cpu to device, set
  batch.capture_hidden_mode, and call ForwardBatch.init_new(batch, runner)
  directly (drop get_model_worker_batch()).

These mirror the equivalent, already-merged handling in eagle3_target_model.py,
so the two backends stay consistent.

Related Issues

Accuracy Test

Benchmark & Profiling

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://sgl-fru7574.slack.com/archives/C09784E3EN6 to discuss your PR.

[gemini-code-assist]

Code Review

This pull request updates the dflash_target_model.py file to maintain compatibility with sglang 0.5.13+. Key changes include bypassing the scheduler's PrefillAdder to manually initialize inputs and extend ranges per request, migrating to the module-level prepare_mlp_sync_batch_raw function, handling host-to-device copying of prefill input IDs, and passing ScheduleBatch directly to ForwardBatch.init_new. Additionally, Req.fill_ids has been replaced with full_untruncated_fill_ids using an array. The reviewer recommends wrapping origin_input_ids in an array("q", ...) to prevent a potential runtime TypeError during execution.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

In sglang 0.5.13+, Req.origin_input_ids is expected to be an array of type 'q' (64-bit signed integers) [5.1.1]. Passing a standard Python list of integers can cause a TypeError during execution, specifically in Req._refresh_fill_ids when concatenating self.origin_input_ids + self.output_ids (since self.output_ids is an array("q") and Python does not allow concatenating a list and an array). Wrapping curr_ids.view(-1).tolist() in array("q", ...) ensures type safety and prevents runtime crashes.

Suggested change

	origin_input_ids=curr_ids.view(-1).tolist(),
	origin_input_ids=array("q", curr_ids.view(-1).tolist()),