[ROCm] mori: add InterNodeV1LL inter-node kernel selection via VLLM_MORI_INTERNODE_KERNEL (#41751)

Signed-off-by: jatseng-ai <jatseng@amd.com>
This commit is contained in:
jatseng-ai
2026-05-28 00:33:32 +08:00
committed by GitHub
parent 41688e2dc7
commit 05c50c721e
7 changed files with 36 additions and 14 deletions
@@ -249,7 +249,7 @@ class Config:
def needs_mori(self):
info = prepare_finalize_info(self.prepare_finalize_type)
return info.backend == "mori"
return info.backend in ("mori_high_throughput", "mori_low_latency")
def all2all_backend(self):
info = prepare_finalize_info(self.prepare_finalize_type)
@@ -232,7 +232,7 @@ if has_mori():
standard_format,
fp8_types,
blocked_quantization_support=True,
backend="mori",
backend="mori_high_throughput",
supports_apply_weight_on_input=False,
)
+5 -3
View File
@@ -92,7 +92,7 @@ PARALLEL_COMBOS = [
BACKENDS = ["allgather_reducescatter"]
if has_mori():
BACKENDS += ["mori"]
BACKENDS += ["mori_high_throughput", "mori_low_latency"]
if has_flashinfer_nvlink_two_sided():
BACKENDS += ["flashinfer_nvlink_two_sided"]
@@ -118,7 +118,8 @@ QUANT_METHODS = [
# fmt: off
BACKEND_SUPPORTED_QUANTS: dict[str, set[str | None]] = {
"allgather_reducescatter": {None, "fp8", "modelopt_fp8", "modelopt_fp4"}, # noqa: E501
"mori": {None, "fp8", "modelopt_fp8"},
"mori_high_throughput": {None, "fp8", "modelopt_fp8"},
"mori_low_latency": {None, "fp8", "modelopt_fp8"},
"flashinfer_nvlink_two_sided": {None, "fp8_blocked", "modelopt_fp4"}, # noqa: E501
"flashinfer_nvlink_one_sided": {None, "modelopt_fp4"}, # noqa: E501
"deepep_low_latency": {None, "fp8_blocked", "modelopt_fp4"}, # noqa: E501
@@ -129,7 +130,8 @@ BACKEND_SUPPORTED_QUANTS: dict[str, set[str | None]] = {
# Map from backend -> (DP/EP support, DP support, TP support, SP support)
BACKEND_EP_DP_TP_SUPPORT: dict[str, tuple[bool, bool, bool, bool]] = {
"allgather_reducescatter": (True, True, True, True),
"mori": (True, False, False, True),
"mori_high_throughput": (True, False, False, True),
"mori_low_latency": (True, False, False, True),
"flashinfer_nvlink_two_sided": (False, True, False, False),
"flashinfer_nvlink_one_sided": (False, True, False, False),
"deepep_low_latency": (True, False, False, True),