[AMD] Add DeepSeek-V4-Pro FP4 MI355X SGLang MTP recipe

.github/configs/amd-master.yaml

@@ -2166,6 +2166,42 @@ dsv4-fp4-mi355x-sglang:
       - { tp: 8, dp-attn: true, conc-start: 64, conc-end: 2048 }
       - { tp: 8, dp-attn: false, conc-start: 1, conc-end: 32 }
 
+# MTP variant of dsv4-fp4-mi355x-sglang. Mirrors the base search space and adds
+# spec-decoding: mtp, which routes to dsv4_fp4_mi355x_sglang_mtp.sh (EAGLE
+# speculative decoding), per sgl-project/sglang#26383 ([AMD][DSV4] DSV4 MTP
+# graph + sparse triton attn optimizations, merged to main 2026-05-27). That PR
+# fixes the ROCm HIP-radix MTP CUDA-graph bug (the false-EOS symptom in sgl
+# #20404) and validates GSM8K 0.950 with MTP on.
+#
+# BLOCKED on an image: no published build both loads DSv4-Pro AND has #26383.
+#   - rocm/sgl-dev:...-DSv4 builds are the only lineage bundling deep_gemm (so
+#     the only ones that load DSv4-Pro), but they ended at f96ac98 (2026-05-27),
+#     pre-#26383 -> MTP graph capture crashes (kv_score dtype, run 26723126211).
+#   - mainline v0.5.12.post1 ROCm nightlies carry #26383 but omit deep_gemm, so
+#     DSv4-Pro weight load fails in _setup_fp8_wo_a_scales (run 26727984372).
+# Pinned to the latest -DSv4 build as the correct lineage placeholder. Unblock by
+# bumping to the first -DSv4 image that carries #26383. PR #1631 parked as draft.
+dsv4-fp4-mi355x-sglang-mtp:
+  image: rocm/sgl-dev:rocm720-mi35x-f96ac98-20260527-DSv4
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: mi355x
+  precision: fp4
+  framework: sglang
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 8, dp-attn: true, conc-start: 64, conc-end: 2048, spec-decoding: mtp }
+      - { tp: 8, dp-attn: false, conc-start: 1, conc-end: 32, spec-decoding: mtp }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, dp-attn: true, conc-start: 64, conc-end: 2048, spec-decoding: mtp }
+      - { tp: 8, dp-attn: false, conc-start: 1, conc-end: 32, spec-decoding: mtp }
+
 # DSv4 on MI355X via vLLM, using the official vllm/vllm-openai-rocm
 # nightly image. DSv4 base ROCm support (vllm-project/vllm#40871) merged
 # on 2026-05-05, so any nightly built after that includes the

benchmarks/single_node/dsv4_fp4_mi355x_sglang_mtp.sh

@@ -0,0 +1,209 @@
+#!/usr/bin/env bash
+
+# DeepSeek-V4-Pro on MI355X via SGLang — MTP variant of dsv4_fp4_mi355x_sglang.sh.
+# Adds EAGLE/MTP speculative decoding per sgl-project/sglang#26383
+# ([AMD][DSV4] DSV4 MTP graph + sparse triton attn optimizations, merged
+# 2026-05-27, commit deaba74), which fixes the ROCm HIP-radix backend's
+# per-step draft out_cache_loc slicing under CUDA graph (the bug behind the
+# false-EOS / truncated-generation symptom in sgl issue #20404) and validates
+# GSM8K 0.950 with MTP on. The EAGLE chain follows that PR's accuracy config
+# for the DP-attention path (steps=2, topk=1, draft=3); the TP-only
+# low-concurrency path uses the (3,1,4) chain shared with dsr1_fp4_mi355x_mtp.sh.
+#
+# IMPORTANT (blocked on image): MTP needs a build that BOTH loads DSv4-Pro AND
+# carries sgl#26383. None exists yet:
+#   - rocm/sgl-dev:...-DSv4 builds are the only lineage bundling deep_gemm (so
+#     the only ones that load DSv4-Pro), but ended at f96ac98 (2026-05-27,
+#     pre-#26383) and crash at MTP graph capture (kv_score dtype, run 26723126211).
+#   - mainline v0.5.12.post1 nightlies carry #26383 but omit deep_gemm and fail
+#     DSv4-Pro weight load in _setup_fp8_wo_a_scales (run 26727984372).
+# amd-master.yaml pins the latest -DSv4 build; bump to the first -DSv4 image with
+# #26383 to unblock. RUN_EVAL on the high-conc points then gates accuracy.
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    DP_ATTENTION \
+    EP_SIZE \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME \
+    MAX_MODEL_LEN
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [[ "$MODEL" != /* ]]; then hf download "$MODEL"; fi
+
+# sglang ships in the image at the SHA encoded in the image tag (built
+# from the amd/deepseek_v4 branch in sgl-project/sglang). To bump sglang,
+# bump the image tag in .github/configs/amd-master.yaml.
+
+# Transformers in the container doesn't recognize the `deepseek_v4` model_type.
+# PR #23608's fallback in hf_transformers_utils.get_config tries to handle this
+# by writing a patched config to /tmp, but in practice isn't catching the error
+# in this image. Patch the cached config.json directly instead: set model_type
+# to `deepseek_v3` so AutoConfig.from_pretrained succeeds, and keep
+# architectures=['DeepseekV4ForCausalLM'] so SGLang dispatches to its native
+# DSv4 model class (python/sglang/srt/models/deepseek_v4.py).
+python3 << PYEOF
+import json
+from huggingface_hub import hf_hub_download
+path = hf_hub_download(repo_id="$MODEL", filename="config.json")
+with open(path) as f:
+    config = json.load(f)
+if config.get("model_type") == "deepseek_v4":
+    config["model_type"] = "deepseek_v3"
+    with open(path, "w") as f:
+        json.dump(config, f, indent=2)
+    print(f"Patched {path}: model_type deepseek_v4 -> deepseek_v3")
+else:
+    print(f"No patch needed: model_type is {config.get('model_type')!r}")
+PYEOF
+
+# DSv4 FP4-experts path. Tracks the env block in python/run_dsv4.sh on the
+# amd/deepseek_v4 branch (HEAD's active block is FP8; we override the two
+# FP4-specific flags below):
+#   SGLANG_DSV4_FP4_EXPERTS=True   -> route experts through the FP4 kernels
+#   SGLANG_FORCE_TRITON_MOE_FP8=0  -> dispatch MoE through aiter and apply
+#                                    the swiglu_limit clamp in the triton
+#                                    MoE fallback path.
+export SGLANG_REASONING_EFFORT=max
+export SGLANG_OPT_USE_FUSED_COMPRESS=true
+export SGLANG_OPT_USE_OLD_COMPRESSOR=false
+export SGLANG_OPT_USE_TILELANG_SWA_PREPARE=false
+export SGLANG_OPT_USE_JIT_KERNEL_FUSED_TOPK=false
+export SGLANG_OPT_USE_FUSED_HASH_TOPK=true
+export SGLANG_OPT_DEEPGEMM_HC_PRENORM=false
+export SGLANG_OPT_USE_TILELANG_MHC_PRE=false
+export SGLANG_OPT_USE_TILELANG_MHC_POST=false
+export SGLANG_OPT_USE_AITER_MHC_PRE=true
+export SGLANG_OPT_USE_AITER_MHC_POST=true
+export SGLANG_ENABLE_THINKING=1
+export SGLANG_USE_AITER=1
+export SGLANG_USE_ROCM700A=1
+export SGLANG_TOPK_TRANSFORM_512_TORCH=0
+export SGLANG_FP8_PAGED_MQA_LOGITS_TORCH=1
+export SGLANG_DSV4_FP4_EXPERTS=True
+export SGLANG_OPT_DPSK_V4_RADIX=1
+export SGLANG_OPT_USE_OVERLAP_STORE_CACHE=false
+export SGLANG_OPT_USE_FUSED_STORE_CACHE=true
+export SGLANG_FORCE_TRITON_MOE_FP8=0
+export SGLANG_HACK_FLASHMLA_BACKEND=triton
+export SGLANG_OPT_USE_TILELANG_INDEXER=true
+export SGLANG_OPT_USE_TRITON_SWA_PREPARE=true
+export AITER_BF16_FP8_MOE_BOUND=0
+export SGLANG_OPT_FUSE_WQA_WKV=true
+export SGLANG_OPT_USE_FUSED_PAGED_COMPRESS=true
+export SGLANG_OPT_USE_MULTI_STREAM_OVERLAP=0
+
+# MTP-specific knobs landed alongside the graph fix in sgl#26383:
+#   SGLANG_OPT_USE_TRITON_FUSED_MHC -> fused Triton mhc_post_pre for low conc
+#                                      (defaults True in post-#26383 images;
+#                                      set explicitly so the recipe is auditable)
+#   SGLANG_OPT_C4_SPARSE_TOPK       -> sparse-attention top-k used in the PR's
+#                                      DSv4 MTP accuracy run
+export SGLANG_OPT_USE_TRITON_FUSED_MHC=1
+export SGLANG_OPT_C4_SPARSE_TOPK=512
+
+SERVER_LOG=/workspace/server.log
+PORT=${PORT:-8888}
+
+EVAL_CONTEXT_ARGS=""
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    EVAL_CONTEXT_ARGS="--context-length $EVAL_MAX_MODEL_LEN"
+fi
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+
+PARALLEL_ARGS=(
+    --tensor-parallel-size "$TP"
+)
+# EAGLE chain is selected by DP_ATTENTION. The DP-attention path mirrors the
+# sgl#26383 DSv4 ROCm accuracy config (steps=2, topk=1, draft=3); the TP-only
+# low-concurrency fallback uses the longer (3,1,4) chain that low batch sizes
+# benefit from, matching dsr1_fp4_mi355x_mtp.sh.
+SPEC_FLAGS=(
+    --speculative-algorithm EAGLE
+    --speculative-num-steps 3
+    --speculative-eagle-topk 1
+    --speculative-num-draft-tokens 4
+)
+if [ "${DP_ATTENTION}" = "true" ]; then
+    PARALLEL_ARGS+=(
+        --dp "$TP"
+        --enable-dp-attention
+        --enable-prefill-delayer
+    )
+    SPEC_FLAGS=(
+        --speculative-algorithm EAGLE
+        --speculative-num-steps 2
+        --speculative-eagle-topk 1
+        --speculative-num-draft-tokens 3
+    )
+fi
+if [ "${EP_SIZE:-1}" -gt 1 ]; then
+    PARALLEL_ARGS+=(--ep-size "$EP_SIZE")
+fi
+
+set -x
+python3 -m sglang.launch_server \
+    --model-path $MODEL \
+    --host=0.0.0.0 \
+    --port $PORT \
+    "${PARALLEL_ARGS[@]}" \
+    "${SPEC_FLAGS[@]}" \
+    --trust-remote-code \
+    --disable-radix-cache \
+    --attention-backend compressed \
+    --max-running-requests ${CONC} \
+    --mem-fraction-static 0.90 \
+    --swa-full-tokens-ratio 0.15 \
+    --page-size 256 \
+    --context-length $MAX_MODEL_LEN \
+    --chunked-prefill-size 8192 \
+    --disable-shared-experts-fusion \
+    --tool-call-parser deepseekv4 \
+    --reasoning-parser deepseek-v4 \
+    --chat-template "$(dirname "$0")/chat_templates/deepseek_v4_thinking.jinja" \
+    --watchdog-timeout 1800 $EVAL_CONTEXT_ARGS > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+# --dsv4 routes prompts through encoding_dsv4.py, emitting the
+# <bos><User>...<Assistant><think> framing DeepSeek-V4-Pro expects. EAGLE/MTP
+# acceptance silently regresses on raw random tokens, so MTP benchmarks must
+# use chat-formatted inputs (AGENTS.md). The DSv4-Pro tokenizer ships without a
+# jinja chat_template, so plain --use-chat-template would crash; --dsv4 handles
+# the framing directly.
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend vllm \
+    --input-len "$ISL" \
+    --output-len "$OSL" \
+    --random-range-ratio "$RANDOM_RANGE_RATIO" \
+    --num-prompts "$((CONC * 10))" \
+    --max-concurrency "$CONC" \
+    --result-filename "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --dsv4
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x

perf-changelog.yaml

@@ -3342,3 +3342,9 @@
   description:
     - "Update vLLM ROCm image from v0.21.0 to v0.22.0"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1616
+
+- config-keys:
+    - dsv4-fp4-mi355x-sglang-mtp
+  description:
+    - "Add MTP/EAGLE speculative-decoding sibling for dsv4-fp4-mi355x-sglang (model: deepseek-ai/DeepSeek-V4-Pro) on rocm/sgl-dev:rocm720-mi35x-f96ac98-20260527-DSv4, per sgl-project/sglang#26383. Blocked pending a -DSv4 image that carries #26383 (PR draft)."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1631