r12 clean build: flashinfer-python 0.6.17 from PyPI (whl-index path dead); recipe MTP:3 comment clarified (built-in MTP head, not dspark)

docker-images/glm53-jovian-r12/Dockerfile

@@ -48,19 +48,17 @@ RUN rm -f /etc/apt/sources.list.d/*.list /etc/apt/sources.list.d/*.sources \
 FROM apt-base AS torch-base
 ARG PYPI_INDEX_URL
 ARG PYTORCH_INDEX_URL
-ARG FLASHINFER_WHL_INDEX
 ARG TORCH_VERSION TORCHVISION_VERSION TORCHAUDIO_VERSION
 ARG FLASHINFER_VERSION
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
     uv pip install \
       --index-url ${PYPI_INDEX_URL} \
       --extra-index-url ${PYTORCH_INDEX_URL} \
-      --extra-index-url ${FLASHINFER_WHL_INDEX} \
       --index-strategy unsafe-best-match \
       "torch==${TORCH_VERSION}" \
       "torchvision==${TORCHVISION_VERSION}" \
       "torchaudio==${TORCHAUDIO_VERSION}" \
-      "flashinfer==${FLASHINFER_VERSION}"
+      "flashinfer-python==${FLASHINFER_VERSION}"
 
 # The fork's r12 composition: base branch + sequential merge of the open
 # PR heads (ascending). The PR head pseudo-refs are GitHub's stable refs.

docker-images/glm53-jovian-r12/versions.env

@@ -17,9 +17,10 @@ TORCH_VERSION=2.13.0
 TORCHVISION_VERSION=0.28.0
 TORCHAUDIO_VERSION=2.11.0
 
-# FlashInfer — not a pip dep of the fork since 0.6.14; from the wheel index
+# FlashInfer — the fork's cuda.txt uses flashinfer-python==0.6.17 (PyPI);
+# flashinfer-cubin is NOT a build dep (setup.py skips it; the runtime
+# fetches cubins when the package is absent).
 FLASHINFER_VERSION=0.6.17
-FLASHINFER_WHL_INDEX=https://flashinfer.ai/whl/cu130/torch2.13/
 
 # vLLM — local-inference-lab fork, dev/jovian-judgement base + PR heads
 VLLM_REPO=https://github.com/local-inference-lab/vllm.git

recipes/glm53-flash-nvfp4-spark/recipe.toml

@@ -111,7 +111,12 @@ args = [
     # Their proven boot: no flashinfer autotune, no CuTe-DSL warmup
     "--kernel-config", """
 '{"enable_cutedsl_warmup":false,"enable_flashinfer_autotune":false}'""",
-    # MTP-5 draft from the embedded NVFP4 MTP shards (lil.yaml mtp_tokens: 5)
+    # Built-in MTP draft, 3 tokens (the r12-qualified "mtp:3" mode — their
+    # x86 numbers: 254 tok/s CC1, 2.37 accept; the checkpoint embeds up to
+    # 5 via the model-mtp-0000x shards but 3 is the verified depth). This is
+    # the checkpoint's own MTP head (Glm5NextMTPModel), NOT dspark — dspark
+    # is the DeepSeek-V4 family's mechanism; GLM's other r12 mode is
+    # dflash2:7 with a separate trained draft checkpoint.
     # The draft model's vllm_config takes its attention backend from the
     # SPECULATIVE config, not the global flag (eagle/utils.py:
     # _make_eagle_draft_vllm_config) — B12X here mirrors the main model's.