feat: integrate dynamic speculative decoding profiling mode and Qwen 3.6 35B matrix#55
Open
feat: integrate dynamic speculative decoding profiling mode and Qwen 3.6 35B matrix#55
Conversation
Contributor
There was a problem hiding this comment.
Pull request overview
Integrates a new “dynamic speculative decoding” profiling configuration into the existing automated profiling runner, and documents benchmark results for the newly added mlx-community/Qwen3.6-35B-A3B-4bit matrix in repo docs/README.
Changes:
- Add draft-model auto-selection + auto-injected speculative decoding config to
scripts/profiling/profile_runner.py. - Add
mlx-community/Qwen3.6-35B-A3B-4bitto the interactive benchmark model list inrun_benchmark.sh. - Publish Qwen3.6-35B benchmark tables in
README.mdand the profiling results markdown.
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 5 comments.
| File | Description |
|---|---|
| scripts/profiling/profile_runner.py | Auto-injects a speculative decoding config by selecting a draft model based on the main model string. |
| run_benchmark.sh | Adds Qwen3.6-35B to the selectable benchmark models list and removes a duplicate entry. |
| docs/profiling/profiling_results_simbas-MacBook-Pro.md | Adds a new Qwen3.6-35B context/memory profiling section. |
| README.md | Adds a new Qwen3.6-35B performance section and headline benchmark table. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+287
to
+292
| # Speculative Decoding Mode Auto-Injection | ||
| draft_model = get_draft_model(args.model) | ||
| if draft_model: | ||
| draft_name = draft_model.split("/")[-1].split("-")[1] # extract roughly '0.8B' or '1b' | ||
| CONFIGS.append({"name": f"TurboQuant + Speculative ({draft_name})", "flags": ["--turbo-kv", "--draft-model", draft_model]}) | ||
|
|
Comment on lines
+289
to
+291
| if draft_model: | ||
| draft_name = draft_model.split("/")[-1].split("-")[1] # extract roughly '0.8B' or '1b' | ||
| CONFIGS.append({"name": f"TurboQuant + Speculative ({draft_name})", "flags": ["--turbo-kv", "--draft-model", draft_model]}) |
Comment on lines
+5
to
+21
| | Configuration | Context Size | TTFT | Generation Speed | Model Size | Active RAM (Physical) | GPU Memory Allocated | | ||
| |---|---|---|---|---|---|---| | ||
| | Dense/Vanilla | 512 | 4.01s | 32.10 tok/s | N/A | 18.9 GB | 33.6 GB | | ||
| | Dense/Vanilla | 40000 | 26.41s | 23.99 tok/s | N/A | 49.4 GB | 64.2 GB | | ||
| | Dense/Vanilla | 100000 | 151.76s | 18.64 tok/s | N/A | 49.3 GB | 63.9 GB | | ||
| | SSD Stream | 512 | 1.81s | 15.01 tok/s | N/A | 4.5 GB | 18.8 GB | | ||
| | SSD Stream | 40000 | 28.89s | 5.13 tok/s | N/A | 37.4 GB | 51.7 GB | | ||
| | SSD Stream | 100000 | 100.72s | 4.08 tok/s | N/A | 49.4 GB | 63.9 GB | | ||
| | TurboQuant | 512 | 0.44s | 33.14 tok/s | N/A | 18.9 GB | 33.3 GB | | ||
| | TurboQuant | 40000 | 20.90s | 2.54 tok/s | N/A | 22.7 GB | 37.0 GB | | ||
| | TurboQuant | 100000 | 60.30s | 4.73 tok/s | N/A | 27.7 GB | 42.0 GB | | ||
| | SSD + TurboQuant | 512 | 1.64s | 14.51 tok/s | N/A | 4.5 GB | 19.3 GB | | ||
| | SSD + TurboQuant | 40000 | 27.56s | 5.39 tok/s | N/A | 8.5 GB | 23.2 GB | | ||
| | SSD + TurboQuant | 100000 | 75.59s | 3.86 tok/s | N/A | 13.6 GB | 28.3 GB | | ||
| | SSD + 16-Worker Prefetch | 512 | 0.94s | 16.70 tok/s | N/A | 4.5 GB | 19.4 GB | | ||
| | SSD + 16-Worker Prefetch | 40000 | 28.88s | 5.17 tok/s | N/A | 37.4 GB | 51.9 GB | | ||
| | SSD + 16-Worker Prefetch | 100000 | 101.96s | 3.79 tok/s | N/A | 49.4 GB | 63.9 GB | |
Comment on lines
+25
to
+26
| if "gemma" in m: | ||
| return "mlx-community/gemma-3-1b-it-4bit" |
| draft_model = get_draft_model(args.model) | ||
| if draft_model: | ||
| draft_name = draft_model.split("/")[-1].split("-")[1] # extract roughly '0.8B' or '1b' | ||
| CONFIGS.append({"name": f"TurboQuant + Speculative ({draft_name})", "flags": ["--turbo-kv", "--draft-model", draft_model]}) |
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.
Integrates the Qwen 3.6 35B deep dive benchmark matrix natively into the SwiftLM README and auto-wires the background speculative decoding validation mode directly into the automated profile evaluation suite to capture multi-model draft performance.