feat(dflash): config-driven drafter converter + spec-decode fixes + rope_parameters Q4 converter

[dusterbloom]

Wave-2 base. The config-driven drafter converter + spec-decode perf/correctness fixes, plus a new corrected converter.

New in this PR (commit 166a409): server/scripts/convert_modal_dflash_to_gguf.py — resolves rope_theta from the nested rope_parameters scope then top-level (legacy), and FAILS LOUD (sys.exit) on a missing theta instead of silently baking a 1M default. This fixes the root cause of the drafter RoPE mismatch: Qwen3.6 drafters publish their true 10M under rope_parameters, which the old top-level-only pick() missed, baking 1M and collapsing long-ctx spec-decode acceptance (gemma4 is genuinely 1M so the bug was invisible there). Q4_K_M output, 12 pure-python tests. Original convert_dflash_to_gguf.py untouched.

Depends on the pooled-prefill core (#430).

[cubic-dev-ai]

8 issues found across 20 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="server/scripts/quantize_draft_q8.py">

<violation number="1" location="server/scripts/quantize_draft_q8.py:30">
P3: `is_norm_tensor` is imported and then redefined locally, so the import is dead/shadowed and adds avoidable duplication.</violation>
</file>

<file name="server/src/qwen35/qwen35_backend.cpp">

<violation number="1" location="server/src/qwen35/qwen35_backend.cpp:191">
P1: Early drafter config sync can copy uninitialized capture layer IDs when `n_target_layers` exists but `target_layer_ids` is missing/invalid.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P1: Early drafter config sync can copy uninitialized capture layer IDs when n_target_layers exists but target_layer_ids is missing/invalid.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At server/src/qwen35/qwen35_backend.cpp, line 191:

<comment>Early drafter config sync can copy uninitialized capture layer IDs when `n_target_layers` exists but `target_layer_ids` is missing/invalid.</comment>

<file context>
@@ -180,6 +180,24 @@ bool Qwen35Backend::init() {
+        std::string dp(cfg_.draft_path);
+        if (dp.size() >= 5 && dp.substr(dp.size() - 5) == ".gguf") {
+            int n_cap = 0;
+            int cap_ids[DFLASH_MAX_CAPTURE_LAYERS];
+            if (read_draft_capture_config(cfg_.draft_path, n_cap, cap_ids,
+                                          DFLASH_MAX_CAPTURE_LAYERS)) {
</file context>

[cubic-dev-ai]

P3: is_norm_tensor is imported and then redefined locally, so the import is dead/shadowed and adds avoidable duplication.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At server/scripts/quantize_draft_q8.py, line 30:

<comment>`is_norm_tensor` is imported and then redefined locally, so the import is dead/shadowed and adds avoidable duplication.</comment>

<file context>
@@ -16,84 +18,20 @@
+# Share arch resolution, tensor-name mapping, and safetensors I/O with
+# the F16 converter so both produce structurally identical GGUF metadata.
+from convert_dflash_to_gguf import (
+    ARCH, load_arch, map_name, is_norm_tensor,
+    load_safetensors_header,
+)
</file context>

Suggested change

	ARCH, load_arch, map_name, is_norm_tensor,
	ARCH, load_arch, map_name,